[FEA] cuda.core: support NVRTC bundled headers (CUDA 13.3) #2363

Summary

NVRTC 13.3 ships bundled CUDA headers plus APIs to manage them — a big usability win for
pip-wheel-only environments that have no CTK installation to point include_path at:

Adjacent: the new NVRTC_ERROR_BUSY (13.3) is plausibly tied to concurrent installs given the
NO_WAIT flag — to be confirmed during design; error mapping itself is automatic.

Interacts with the cuda-cccl dependency discussion (#691) and cuda.pathfinder's include-dir
discovery.

Design sketch (draft — needs design-meeting review)

• ProgramOptions.use_bundled_headers: str | bool (a True shorthand meaning "the managed
  default install location").
• cuda.core.utils.install_bundled_headers(path=None, *, policy=...),
  remove_bundled_headers(path), bundled_headers_info() wrappers.
• Optionally: Program auto-installs (opt-in) to a per-user cache dir on first use.

Open questions for the meeting:

1. Who owns the default install location (per-user cache dir? env var? pathfinder-managed)?
2. Concurrency/NO_WAIT policy and the NVRTC_ERROR_BUSY UX.
3. Interplay with include_path, no_source_include, and pathfinder-discovered CTK headers.
4. Should Program ever auto-trigger installs, or keep it strictly explicit?

References

• NVRTC docs: https://docs.nvidia.com/cuda/nvrtc/index.html
• Found during the CUDA 12.8 → 13.3 bindings vs. cuda.core gap sweep (2026-07-14)

-- Leo's bot