Flocking is now built from the generic ops, so the built-in flock goes.

• Removed: p.flock(), particles_flock, particles_kernel_flock, the flock.wgsl kernel, and its FFI constructor.
• flocking_gpu.py is now a recipe with no shader code: find neighbors, sum what each rule steers toward, apply steering with a mask for boids that have no neighbors, then add the force to velocity and limit the speed.

It's slower than the fused kernel: about 3.5 ms vs 1.6 ms at 20k boids, and 8 ms vs 4 ms at 100k. Most of that is the per-dispatch cost of ~30 passes, not GPU work, so batching submits is the next thing to look at. Anyone who needs the speed can still write a small kernel that reads neighbors directly.