core: a direct read the file system refuses goes through the page cache

[plotarmordev]

What this changes

When a file system accepts an O_DIRECT open but fails the aligned read with EINVAL, core.direct_io now clears
O_DIRECT on that descriptor and reads through the page cache instead of failing the load (part of #914).
Nemotron's CUDA loader already reads through direct_io.

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision:
  • Commit 6bb9daaca25a, on main c7137bb3. No MLX or PyTorch.
  • An NVIDIA GB10 build host on Ubuntu 24.04.4 LTS aarch64 (Linux 6.17), with Zig 0.17.0 and nvcc 13.0.88.
  • No checkpoint is involved.
• Exactness, from tools/bench_concurrent.py --alone --serial or token_sha against "draft": false: not run. No
  token path is touched. The new test checks that the bytes read are the file's bytes.
• Decode speed before and after, from tools/bench_openai.py: not run. Decode does no file reads.
• Prompt speed before and after: the cold first request's ttft_s from tools/bench_cache.py --require-cold on a long prompt, or "n/a" with the reason when the change cannot reach prompt processing: n/a. Only checkpoint
  loading reads files, and loading on file systems that serve O_DIRECT reads takes the unchanged path.
• Tests run, with pass, fail and skip counts:
  • CUDA_VISIBLE_DEVICES= zig build test-cuda-host --summary all from an empty local cache: 27/27 steps, 466 passed, 0 failed, 0 skipped.
  • The new test, "a direct read the file system refuses is read again through the page cache", injects EINVAL
    through the reader's pread hook.
  • The same test added to main's reader fails with ReadFailed ("read of 102400 bytes at 12288 failed: INVAL").
    Command: CUDA_VISIBLE_DEVICES= zig build test-cuda-host on c7137bb3 plus only this test and its Refuse
    helper.
  • zig build native -Dnvcc=/usr/local/cuda/bin/nvcc --summary all: 42/42 steps.
  • python3 tools/zig/lean_check.py: 527 problems, the same as main's 527 (tools/zig/lean_new.py: 0 new).
  • zig fmt --check build.zig zig: clean.
• Not run, and why:
  • Metal: no macOS host here. The reader is the Linux path, and the new test skips on other systems.
  • A file system that actually refuses O_DIRECT reads was not available, so the refusal is injected.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (No token
  path is touched.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Not
  run; no output path is touched.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (No
  arithmetic changes.)
• Prompt processing is no slower than the last release. (Not measured; loading only.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (It skips off
  Linux or without O_DIRECT.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.