[CanReader]

• batch tensor I/O into single read/write calls in checkpoint serialization (by @sharmis-tp, from Batch tensor I/O in checkpoint serialization #76)
• format

What this changes

save_tensors converts each tensor to bytes once and writes it with one call; load_tensors reads a tensor's bytes with one call and decodes them. The on-disk format is unchanged.

Why

This is #76 rebased across #74, which moved the body of save_tensors into the atomic-write closure and left #76 conflicting. The contributor's commit is kept as is with their authorship; the second commit is cargo fmt. Landing it from a branch in this repository because the fork's branch cannot be updated from here.

Closes #54.

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

Serialize unit tests, robustness and safetensors suites pass locally on the rebased result; CI runs the rest.

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