cuda.core: add numba-cuda-mlir port of the Stable Fluids example

Add examples/gl_interop_fluid_numba_cuda_mlir.py, a numba-cuda port of
gl_interop_fluid.py. numba has no texture/surface support, so the same
Stable Fluids solver runs on plain linear device arrays with a hand-written
bilinear sampler; the physics and pipeline are unchanged from the CUDA C++
version, only the memory model and read path differ.

It targets the MLIR numba-cuda backend (numba-cuda-mlir), which tracks the
current cuda.bindings/cuda.core API. Classic numba-cuda lags the editable dev
tip (cuda.core.graph relocation per #1858, property-style kernel attributes),
so it cannot launch kernels against it.

Declare numba-cuda-mlir as a linux-gated pypi dependency of the test feature
so `pixi run python examples/...` works in the default env. Upstream ships
only manylinux cp311-cp314 wheels (no win-64), hence the linux target gating.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: rparolin