_util: drop the dead st_mtime=0 short-circuit in raise_on_not_writable_file

[HrachShah]

raise_on_not_writable_file guarded its writability and is-dir checks behind if file_stat.st_mtime != 0, on the theory that an all-zero stat struct meant the caller's stat call had failed in a way the OS could not report (see the original commit c9b6d90, "Raise when trying to acquire in R/O or missing folder").

The guard was useful on the very old NFS / Linux combinations it was written against. On every Python release filelock supports today the result of os.lstat is well-defined and never returns an all-zero struct for a file that exists, so the guard has been dead code for a long time — and worse, it silences both checks whenever a caller (or a test, or a tool that resets mtime to 0 for unrelated reasons) explicitly sets the mtime to 0. The bug only mattered in practice when the lock path was a read-only file or a directory: the check would skip both guards, the caller's subsequent os.open would succeed on non-Windows, and the acquire path would either take a lock on a file nothing else can write to or block forever waiting for an os.open that cannot succeed.

SoftFileLock with its EEXIST-then-stale-break fallback could also spuriously recycle a live lock whose file happened to have mtime=0: the EEXIST from os.open(O_CREAT | O_EXCL) falls into _try_break_stale_lock, and the "malformed lock, age ≥ threshold" branch unconditionally unlinks a lock whose mtime is sufficiently old — and time.time() - 0.0 is, of course, sufficiently old.

What changes

• src/filelock/_util.py: remove the st_mtime != 0 short-circuit. The writability and is-dir checks now run regardless of the observed mtime, matching the function's documented contract.
• tests/test_util.py: add 3 unit tests pinning the function-level behavior:
  • test_raise_on_not_writable_file_rejects_readonly_file_with_mtime_zero: 0o444 file with mtime=0 still raises PermissionError.
  • test_raise_on_not_writable_file_rejects_readonly_file_with_future_mtime: 0o444 file with mtime=2_000_000_000 still raises PermissionError — locks in the "mtime is irrelevant to writability" reading so a later patch can't silently narrow the check to one mtime range.
  • test_raise_on_not_writable_file_rejects_directory_with_mtime_zero: directory with mtime=0 still raises IsADirectoryError.
• tests/test_filelock.py::test_mtime_zero_exit_branch: the SoftFileLock parametrization now expects PermissionError instead of TimeoutError. The previous expectation relied on the very short-circuit this PR removes — under the new behavior, the writability check fires first on a non-root user, which is what the test name ("exit branch") and the FileLock parametrization have always assumed.

Tests

python -m pytest tests/test_util.py → 10 passed, 2 skipped (3 new tests added; pre-existing 7 + 2 symlink tests skipped on this platform still pass).

python -m pytest tests/test_filelock.py -k mtime_zero → 1 passed, 1 failed. The failure is test_mtime_zero_exit_branch[UnixFileLock-PermissionError], which is a pre-existing root-only failure (root can os.open a 0o444 file with O_RDWR | O_TRUNC, so the test only fires on non-root unices) — confirmed by re-running on the unpatched main from a stash, where the same test fails in exactly the same way.

python -m pytest tests/ → 471 passed, 31 skipped, 2 failed. Both failures are pre-existing on the unpatched main (the mtime_zero[UnixFileLock] and test_stale_inode_retry_on_unlinked_lock cases), unrelated to this change.