chmod: report the real error when a target's metadata is inaccessible

[0xfandom]

Fixes #9789.

chmod used Path::exists() to decide whether a target exists. Path::exists() returns false for any metadata access error, including permission errors, so a file whose parent directory lacks search permission was reported as:

chmod: cannot access 'locked/file': No such file or directory

instead of GNU's:

chmod: cannot access 'locked/file': Permission denied

This switches the non-recursive path to Path::try_exists(), which returns Ok(false) only for a genuine ENOENT and Err(..) for a permission error. A permission error is now reported via the existing ChmodError::PermissionDenied variant, whose message already matches GNU. Any other unexpected metadata error falls through to the normal chmod attempt so it can surface a precise message.

Added a regression test (test_chmod_inaccessible_file_reports_permission_denied) for the non-recursive case; it fails on the old Path::exists() code (prints "No such file or directory") and passes with the fix. The existing permission-denied tests only covered the recursive (-R) path.

[github-actions]

GNU testsuite comparison:

Note: The gnu test tests/seq/seq-epipe is now being skipped but was previously passing.

[sylvestre]

a few jobs are failing

[codspeed-hq]

Merging this PR will improve performance by 22.62%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 1 improved benchmark
✅ 344 untouched benchmarks
⏩ 46 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	df_with_path	700.8 µs	571.5 µs	+22.62%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing 0xfandom:chmod-inaccessible-permission-denied (6dadcce) with main (a730551)}

Footnotes

1. 46 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[0xfandom]

Thanks for the heads up. The failing job was clippy's unreadable_literal on the new test — I'd written the octal mode as 0o100644. Pushed a fixup that changes it to 0o100_644, and clippy is clean locally now.

The CodSpeed report flags a ~3.5% change on du_wide_tree, but that's an unrelated du benchmark this PR doesn't touch, and the report itself notes it compared across different runtime environments. Should be noise. Let me know if you'd like anything else adjusted.