Use pathlib in cuda.pathfinder._headers (part 8 of #2410)

[LeSingh1]

Part 8 of #2410 (os.path -> pathlib.Path), covering cuda_pathfinder/cuda/pathfinder/_headers.

This is the last un-migrated corner of cuda.pathfinder runtime code that isn't already
covered by an open part of the series.

What changed

• _locate_in_anchor_layout() and resolve_conda_anchor() build and return Path
  objects, and their anchor arguments accept str or any os.PathLike[str] (per your
  note on Migrate from os.path to pathlib.Path #2410 about widening what we accept).
• The os.path.isfile(os.path.join(...)) probes become (dir / name).is_file(), which
  makes the _joined_isfile helper unnecessary.
• LocatedHeaderDir.abs_path stays a str: it is public API (LocatedHeaderDir is
  re-exported from cuda.pathfinder), so the find steps convert at the boundary, the
  same way Use pathlib in cuda.pathfinder._static_libs (part 2 of #2410) #2493 does for LocatedBitcodeLib/LocatedStaticLib.

Two deliberate exceptions, flagged for review

1. _abs_norm() keeps os.path.abspath(). pathlib has no purely lexical
   normalization: Path.resolve() follows symlinks, which would change the reported
   header directory under the very common /usr/local/cuda -> /usr/local/cuda-<ver>
   layout, and Path.absolute() does not collapse ... I'd rather leave one clearly
   commented os.path call here than silently change what the public API returns.

   While there, the wrapping os.path.normpath() is dropped as redundant: both
   posixpath.abspath and ntpath.abspath already normalize.

2. glob.glob() stays for the descriptor catalog's glob patterns. Only the
   os.path parts of those expressions move to Path. Path.glob() is not a drop-in
   replacement (it needs a separate base directory for absolute patterns, and it does
   not skip dot-entries the way glob.glob() does), so swapping it in would be a
   behavior change unrelated to this issue.

Following your call on #2410, the single-file is_file() probes keep pathlib's error
behavior rather than being wrapped to mimic os.path.isfile()'s "any stat error means
False".

Verification

No behavior change is intended, so there is no new test; the existing suite is the
check. cuda_pathfinder/tests/test_find_nvidia_headers.py exercises all five find
steps, including the conda targets/*/include layout and the CTK-root canary fallback.

• pytest cuda_pathfinder/tests — same pass/fail set as upstream/main on this machine
  (the only failures are the ones that need a real CUDA install).
• mypy cuda/pathfinder/_headers/ — no new errors (the two pre-existing
  redundant-cast errors in header_descriptor.py are untouched and out of scope).
• ruff check and ruff format --check clean on both files.

Refs #2410.

[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.