Upgrade both GPU (runtime) and CPU (cpu-runtime) Colab base images to
the latest available release:

- GPU: `release-colab-external-images_20260416-060047_RC00`
- CPU: `release-colab-external-images_20260416-060047_RC00`

Previously using `release-colab-external_20260226-060109_RC00`.

Note: The tag naming convention has changed from
`release-colab-external_*` to `release-colab-external-images_*`.

b/493600019

## Problem

After merging Kaggle#1547 (Colab base image upgrade), the main branch CI is
failing on the **Test CPU Image** stage (build #1924). The upgraded
`numba-cuda` v0.30.0 now depends on `cuda-bindings` which requires
`libcudart.so` at import time — this crashes on the CPU image where no
CUDA runtime is installed.

Two tests fail:
- `test_numba` — `from numba import cuda` at module level triggers the
crash
- `test_tsfresh` — `tsfresh` → `stumpy` → `from numba import cuda` →
same crash

```
cuda.pathfinder._dynamic_libs.load_dl_common.DynamicLibNotFoundError:
Failure finding "libcudart.so": No such file: libcudart.so*
```

## Fix

### Dockerfile.tmpl
- Keep `numba` upgrade for both CPU and GPU images (needed for NumPy
2.4)
- Move `numba-cuda` install into the GPU-only section (`{{ if eq
.Accelerator "gpu" }}`)

### test_numba.py
- Move `from numba import cuda` from module-level into the `@gpu_test`
method (lazy import)

### test_tsfresh.py
- Guard the `tsfresh` import with try/except and skip the test if it
fails on CPU

b/485275559

kornia.utils.image_to_tensor is deprecated and not lazy loaded anymore.
Using the new location instead.

http://b/515825952

yeah also note this will also remove tf decision forrest , since it is
not compatible with tf 2.20
Tf doc mentions users should switch over to ydf, which is in this image.
so we are okay to drop

Add a dependency on `grpcio-tools` so we can recompile proto stubs in
real time for kaggle_evaluation. It's a small library (<3 MB) and for my
purposes the exact version doesn't much matter so I'm hoping this isn't
a controversial addition. We do have alternatives if this addition isn't
feasible but they're much more invasive, most likely migrating from gRPC
to HTTP.

…le#1556)

BUG=[b/545164527](http://b/545164527)

The scheduled build re-resolves kaggle_requirements.txt against live
PyPI, so a rebuild of an unchanged commit picks up new upstream
releases. Two major versions landed since the last green build and broke
all four test stages:

- fury 2.0.0 swapped the VTK backend for pygfx/wgpu, dropping the APIs
dipy.viz needs. test_dipy fails at import.
dipy/dipy#3978
- fastcore v2 removed L.starmap, called by fastai's set_hypers.
test_fastai.test_tabular fails with AttributeError.
fastai/fastai#4154

Both consumers declare unbounded requirements (fury>=0.12.0,
fastcore>=1.14.6), so nothing stopped the resolver from crossing the
major version boundary. Neither upstream has a released fix, so pin both
with TODOs.

fastcore is capped directly rather than via a direct dep: five packages
pull it uncapped, and execnb>=2.2.5 alone forces v2, so pinning nbdev
does not help. Even nbdev<3 still resolves fastcore 2.2.12.

Verified by building the CPU image both ways. Before: fury 2.0.0 /
fastcore 2.2.12, both tests fail with the CI tracebacks. After: fury
0.12.0 / fastcore 1.14.5 / nbdev 3.1.0, both tests pass and nbdev,
execnb, ghapi and learntools still import.

See b/545164527#comment3 for successful build

Co-authored-by: Claude <noreply@anthropic.com>

Bumps the Colab base image from
release-colab-external-images_20260514-060047_RC00
to release-colab-external-images_20260716-060051_RC00.

Package versions picked up from the new base include torch 2.10 -> 2.11,
torchvision 0.25 -> 0.26, gradio 5.50 -> 6.20, langsmith 0.7.34 ->
0.10.2,
gitpython 3.1.47 -> 3.1.51, pyjwt 2.12.1 -> 2.13.0,
python-multipart 0.0.26 -> 0.0.32, litellm 1.82.4 -> 1.85.7,
wandb 0.26.1 -> 0.28.0, and rsync 3.2.7-0ubuntu0.22.04.4 -> .22.04.6.

Repo-side changes needed alongside the bump:

- torchcodec 0.10.0 -> 0.11.0, to stay compatible with torch 2.11.
- pillow>=12.2 and urllib3>=2.7, since the new base still ships pillow
11.3
  and urllib3 2.5. Both resolve cleanly with no conflicts.
- Stop installing git-lfs via apt. The base image ships git-lfs 3.7.1
built
with Go 1.26, and Ubuntu's git-lfs 3.0.2 package was overwriting it with
an
  older build from Go 1.18.

Both CPU and GPU images build clean with --no-cache. Test suite not yet
run.

The P100 Jenkins agent (label `ephemeral-linux-gpu`) is being deleted,
so this moves everything that ran on it over to T4x2.

## Changes

- **Removed the `Test on P100` stage.** T4x2 becomes the sole GPU test
bed.
- **Repointed `Build GPU Image` / `Diff GPU Image`** at
`ephemeral-linux-gpu-t4x2`. This part is load-bearing, not cleanup:
`ephemeral-linux-gpu` *is* the P100 agent, so those stages would block
indefinitely on a nonexistent executor once it's torn down.
- **Dropped the dead `p100_exempt` decorator** from `tests/common.py`
and its 10 uses across 8 test files, plus the stale cuDNN/sm_60 comment
blocks. Those tests were only ever skipped on Pascal hardware, so they
now run unconditionally on T4.

### Base image
- GPU and CPU images now use
`release-colab-external-images_20260917-060051_RC00` (previously
`20260716-060051`).
- The new base image moves from Python 3.12 to Python 3.13.

### Dockerfile.tmpl
- **Python paths:** `PACKAGE_PATH` and the `sitecustomize.py`
destination now point to `python3.13` instead of `python3.12`.
- **Stopped installing the `keras_internal.py` patch:**
`tensorflow_decision_forests` is no longer in the base image, and its
latest release (1.12.0) has no cp313 wheel, so the
`tensorflow_decision_forests/keras/` directory it patched doesn't exist.
- **vtk apt dependency:** replaced `libgl1-mesa-glx`, which newer Ubuntu
no longer provides, with `libgl1` and `libglx-mesa0`.

### kaggle_requirements.txt
**Removed**
- `tensorflow-io`: the last release (0.37.1, July 2024) requires
`python<3.13`, has no cp313 wheels, and the project is unmaintained.
- The `jupyter-lsp==1.5.1` pin (b/276358430): on 3.13 it keeps
jupyterlab on 3.x, which pulls in `jupyter-ydoc 0.2.5` and then `y-py
0.6.2`. `y-py` has no cp313 wheel and its sdist no longer builds.
jupyterlab 4.x uses `pycrdt` instead, so jupyter-lsp is now unpinned.
- The `fastcore<2` pin and the `python-fasthtml<0.14.6` pin that only
existed because of it (b/545164527).

**Added or changed**
- `opencv-python<5`, `opencv-python-headless<5` and
`opencv-contrib-python<5`: the base image mixes OpenCV versions
(`opencv-python`/`-headless` at 5.0.0.93, `opencv-contrib-python` at
4.14.0.94). All three install into the same `cv2/` package and overwrite
each other, and OpenCV 5.0 drops long-standing APIs such as
`cv2.CascadeClassifier`. Holding all three on 4.x keeps `cv2` consistent
and backward compatible.
- `flax` is now `flax>=0.12`.
- Added `universal-pathlib` because Pyxis
([Kaggle#1402](Kaggle/kaggle-environments#1402),
merged Sep 18) added `universal-pathlib` to kaggle-environments'
[pyproject.toml](https://github.com/Kaggle/kaggle-environments/blob/d7729da06cc1382eb742d6980dc3180aa85caa28/pyproject.toml#L41).

b/493600019