fix(geometry): accept path-like filenames in PLY point-cloud I/O - #5092

Thanks. This fixes #5072 as the issue proposed: os.fspath first, then the same suffix and missing-file checks. I checked it on main a30540a with torch 2.14.0 (macOS arm64, CPU and MPS). I pushed one small commit on top, 0f7867b, which changes the error message and the changelog wording. Please pull before you push again. Details below; I'm approving with it.

What I verified

• The issue's repro passes. A pathlib.Path and a custom os.PathLike[str] now save and load in all four functions. On main they raised TypeError.
• Rejected inputs still raise TypeError, for every input I tried, in all four functions:
  • 1234, None and a list: the exception type is unchanged, but the message is now Python's own expected str, bytes or os.PathLike object, not int. That's fine.
  • bytes and a bytes PathLike: os.fspath returns bytes, the isinstance(filename, str) check rejects it, and the message is kornia's, as on main.
  • a __fspath__ that returns an int: os.fspath raises TypeError.
  • a Path ending in .txt.
  • A Path naming a missing .ply file still raises ValueError in the loaders.
• .PLY is not a behaviour change. Main already accepted "x.PLY" as a str (.lower()), and the PR keeps that.
• The output bytes are unchanged. For the same points, the ASCII and binary files have the same SHA-256 on main and here. A file saved through a Path or a custom PathLike is byte-identical to the same save through a str. The writers are untouched, so nothing new happens with line endings on Windows. Your byte comparison checks two outputs of the same writer, and the Windows leg passes.
• Tests. Your new class fails on main (52 failed: 48 test_pathlike_filename_5072 and 4 test_missing_file) and passes here. The two tests you removed from TestSaveLoadPointCloud (1234 and "pointcloud.txt" to the ASCII saver) are covered by the new parametrized ones, which run all four functions. The file passes on CPU float32/float64, float16/bfloat16 under their known-failure profiles, and MPS float32. The half-precision RNG audit passes too.
• Mutants. Each of the four new os.fspath lines, when dropped, fails 12–14 tests, and str(filename) in its place fails 6–7. Dropping .lower() or the isfile check fails 6–42 tests. Dropping isinstance(filename, str) survives, but the mutant is equivalent: after os.fspath it only sees bytes, and bytes.lower().endswith(".ply") raises TypeError anyway.

What I changed (0f7867b)

• Error message. It said Input filename must be a string with the .ply extension, so save_pointcloud_ply(Path("points.txt"), ...) told the caller to pass a string. It now reads Input filename must be a str or os.PathLike[str] with the .ply extension, in both savers and _check_ply_filename. test_invalid_filename_extension matches that wording instead of "extension".
• Changelog. The fragment now says these filenames raised TypeError before. It also says the case-insensitive .ply check is the one strings already had, so "including case-insensitive .ply extensions" doesn't read as a new feature.

Our docs PR #5069

#5069 documents this defect, so it conflicts in kornia/geometry/pointcloud.py. This PR merges first, and we'll update #5069 afterwards; nothing for you to do.

Posted on behalf of @ducha-aiki by Claude (Opus 5.5).