[CanReader]

• one atomic write helper for checkpoints and the dataset cache

What this changes

serialize::atomic::write_atomic is the only atomic write now. It gains what the copy in mnist.rs had: a per-call counter in the temp name next to the pid, and cleanup of the temp file on any failure, the rename included. write_atomic_bytes covers callers with the bytes in memory, and mnist.rs uses it instead of its own helper. The helper's tests move into atomic.rs and now cover the two error paths: a write that fails midway leaves the old file byte-identical and no temp behind, and a failed rename leaves no temp. In tests/robustness.rs the two "does not destroy the good copy" tests are renamed to what they check, saving over a corrupt file, the test that asserted the temp is absent after a successful save is gone since the rename guarantees that, and the section divider is gone. The two identical doc paragraphs on the savers are one line each.

Why

#72 and #74 each added an atomic write, in different modules with different gaps. One helper, with both sets of behaviour and tests for the paths that matter, is what the reviews on both asked for.

How it was verified

• cargo test passes
• cargo clippy --all-targets -- -D warnings is clean
• Ran cuda_parity on a GPU
• Trained an example in --release
• No GPU here, CUDA side not runtime tested

cargo doc with -D warnings is clean. No CUDA code touched.

Checklist

• Every new or changed backward rule has a gradcheck case
• Every new kernel has a cuda_parity case
• tensor/ops and autograd/ops still mirror each other
• Public items have /// docs