Keep CPU installations platform-neutral while making CUDA runtime and buffer dependencies explicit and deterministic.

Select the runtime provider at source-build time and expose one C++ and Python conversion API across CPU and optional CUDA backends.

Coordinate the Python runtime with the native provider variant and group both conversion packages under one source tree.

Remove the hidden Python stream fallback so applications control stream ownership and consistently share it with ONNX Runtime.

Defer optional CUDA-buffer integration to consumer configuration so the conversion package remains a header-only portable interface.

Rename the conversion adapters to plugins, move the shared API into the
onnxruntime_conversions facade package, and add mutually exclusive CPU and
CUDA runtime packages for C++ and Python. The CUDA plugin now serves both
cpu and cuda backends, so the CUDA runtime no longer ships a second plugin
class.

…k core

onnxruntime_conversions becomes a header-only adapter over
dlpack_conversions, and onnxruntime_conversions_py becomes pure Python
over dlpack_conversions_py. Storage no longer arrives through
ONNX-specific plugins, so the per-device packages are gone:
onnxruntime_conversions_cpu, onnxruntime_conversions_cuda,
onnxruntime_conversions_py_core, onnxruntime_conversions_py_cpu and
onnxruntime_conversions_py_cuda are removed, and their tests move into
the two remaining packages.

Execution provider selection moves into the adapter, where
configure_session_options and session_providers name a provider for the
backend that allocated the storage. Because the adapter compiles in the
consumer's translation unit, an ONNX Runtime upgrade no longer requires
rebuilding the storage plugins, and CPU, CUDA and ROCm storage can be
installed and chosen at runtime.

BREAKING CHANGE: onnxruntime_conversions is now header-only and the
per-device conversion plugin packages no longer exist. Replace
allocate_tensor_msg device arguments with a backend name, and build
session options through configure_session_options or session_providers.

…uildable

The Python to_tensor_msg staged every copy through OrtValue.numpy() and
update_inplace, neither of which works on external device memory, so any
device destination failed outright. Route it through the DLPack core's new
copy path instead, which hands the copy to the storage plugin that owns
the memory.

The pub/sub component nodes call the CUDA runtime directly but were built
unconditionally and never linked cudart, so a CPU-only build of this
package could not compile its tests. Build them, the CUDA gtest, and the
launch test only where the toolkit is present.

Also drop the vendor conflicts against packages this branch deletes, drop
the launch test dependencies the pytest rewrite left behind, and reject a
nonzero DLPack byte_offset rather than silently handing ONNX Runtime a
base pointer.

The READMEs still named the per-device plugin packages this refactor
replaced and still passed an Ort::MemoryInfo and a ConversionConfiguration
that the adapter no longer takes, so every documented example was
uncompilable. Describe the storage plugin split and the current
signatures.

Accept stable ONNX Runtime >=1.27.0 with provider and API checks. Pin fallback providers to 1.29.0, select CUDA dependencies by toolkit family, and document source reuse and Debian installation.

Share Identity and MatMul fixtures across native unit and interprocess tests. Construct the graphs with ONNX Runtime instead of embedding serialized byte arrays.

Allow toolkit versions >=13.1,<14 and document metapackage selection and linked runtime dependencies.

Stage ARM64 CUDA libraries from the ONNX Runtime wheel and reuse installed
SDKs using version metadata and CUDA provider checks. Remove the native
compile/run probe, Python tensor-operation checks and CUDA runtime hook.

Declare NumPy for Python vendor builds and document source reuse, fallback
versions and existing-toolkit builds in the conversion README.

Validated vendor selection and tests, CPU Debian consumers, CUDA wheel
staging and GPU import orders on x86. ARM source builds were validated
before the final discovery cleanup.

Restructure Python and C++ examples and document accelerator setup and Debian/source requirements. Remove the ROSIDL_TENSOR_BACKEND override and CUDA upper bound, and configure cppcheck with GoogleTest definitions.