cuda.bindings: add decode_c_str helper for actionable UnicodeDecodeError (#2122)

[aryanputta]

Summary

Adds cuda.bindings._internal.strdecode.decode_c_str(data, api_name).
Success path = bytes.decode() as today. On UnicodeDecodeError, re-raises
the same exception type with reason extended to include the originating
CUDA API and a bounded hex preview of the buffer.

Implements the helper proposed inline on #2118 ("a tiny helper... source
API name plus a bounded repr/hex dump"). #2118 fixed the WSL symptom in
cuda_core; this closes the diagnostics gap in cuda_bindings.

Design

• Pure-Python _internal/strdecode.py next to _fast_enum.py — no Cython rebuild
• Same exception type; original chained via from e
• Hex preview, stops at first NUL with stopped at NUL@<offset> marker,
  caps at 64 bytes with +N more

Example

UnicodeDecodeError: 'utf-8' codec can't decode byte 0xf8 in position 0:
invalid start byte (returned by nvmlSystemGetProcessName;
bytes=<5 bytes; hex='f8 9a 80 80 af'>)

Scope of this PR

Two files, both hand-written:

• cuda_bindings/cuda/bindings/_internal/strdecode.py — the helper
• cuda_bindings/tests/test_strdecode.py — 11 unit tests

The existing _output_.decode() sites in nvml.error_string,
nvvm.get_error_string, nvfatbin.get_error_string, and
cufile.cufileop_status_error are good candidates to adopt this helper.
Those modules are maintained on the NVIDIA side, so wiring is left to the
maintainers rather than included here.

Test

All 11 tests pass locally; ruff lint and format clean.

Refs #2118. Closes #2122.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

Needs-Restricted-Paths-Review was assigned by CI: Restricted Paths Guard.

For details, open this workflow run and click Summary.