Add CUDA 13 runtime headers

[jasp-nerd]

Proposed changes

Fixes #3859.

The CUDA 13 backend wheel does not depend on the package containing
cuda_runtime.h. After that package is installed, its CUDA 13 headers are
under nvidia/cu13/include, while MLX only checks the CUDA 12
nvidia/cuda_runtime/include layout.

This change:

• adds the CUDA 13 runtime package dependency;
• resolves both the CUDA 12 and CUDA 13 package layouts while preserving the
  existing toolkit fallback;
• requires the runtime header pair before accepting an include directory; and
• verifies the unconditional dependency in the built CUDA 13 wheel metadata.

The path tests use isolated package trees so a system CUDA installation cannot
mask a missing package lookup. CCCL is not added because MLX already bundles
the CCCL headers used by its JIT kernels.

Validation:

focused path tests: 1 test case, 5 assertions passed
CPU C++ suite: 245 test cases, 3,243 assertions passed
mocked CUDA 13 package metadata: passed
pre-commit hooks on changed files: passed

This machine does not have a CUDA toolkit or NVIDIA GPU, so the real CUDA 13
wheel build, JIT smoke test, and Windows coverage are left to repository CI.

I used an AI coding assistant to help research and prepare this change; I
remain responsible for the submitted patch.

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (not needed for this packaging fix)