chunkers: only use the direct read path for known regular files

[ThomasWaldmann]

Fixes the deterministic NetBSD CI hang introduced with PR #10060's direct read path.

The hang: test_create_read_special_symlink (FIFO via create --read-special) hung NetBSD 10.1 forever, at the same spot in both attempts of the PR #10060 CI run — one pytest-xdist worker stuck in the test from minute 2, the suite stalled at 88%. The earlier branch push without the direct-read commit passed NetBSD in 39 min, and FreeBSD/OpenBSD/OmniOS/Linux pass with it, so it bisects cleanly to the direct path reading a FIFO. Special files are opened blocking, and NetBSD was seen never returning from the direct path's big readv() on the FIFO, while the buffered block reader's 1 MiB os.read() calls have always worked there.

The fix: only known regular files take the direct path.

• FileReader requires S_ISREG for direct. chunkify() (ChunkerBase, ChunkerFixed, ChunkerFailing + all .pyi stubs) accepts an optional st (os.stat_result) for callers that already have one.
• process_file() passes the stat result it already has — the create hot path adds no stat syscall. Regular files keep the zero-copy direct path; --read-special FIFOs/devices route onto the buffered path via their stat mode.
• Callers without a stat at hand still get the direct path where safe: FileReader fstats fh / fd.fileno() itself (one cheap syscall per file). Objects without an OS-level fd (BytesIO, recreate/transfer chunk-iterator wrappers) stay on the buffered path, as before fixed chunker: read directly into per-chunk buffers #10060.
• New tests: direct path for a regular file with st given (data + EOF verified); direct path via FileReader's own fstat (fh and file object); buffered for file-likes without an OS fd; direct refused for a FIFO's st.

Local runs: chunkers suite 114 passed, create_cmd + benchmark tests 72 passed, mypy unchanged (same 12 pre-existing unrelated errors).

Note: with PR #10064's PYTEST_TIMEOUT=300, a hang like this would fail in 5 minutes with a stack dump instead of stalling CI for hours — the two PRs complement each other.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.76%. Comparing base (f9e6756) to head (800a4b8).
⚠️ Report is 6 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #10065   +/-   ##
=======================================
  Coverage   86.75%   86.76%           
=======================================
  Files          98       98           
  Lines       17098    17100    +2     
  Branches     2590     2589    -1     
=======================================
+ Hits        14834    14837    +3     
  Misses       1570     1570           
+ Partials      694      693    -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.