trainer.ModelCheckpoint cannot save: uses msgpack_from_state_dict (a restorer) as a file writer

[chaoming0625]

ModelCheckpoint._save_checkpoint calls braintools.file.msgpack_from_state_dict(checkpoint, filepath), but that function restores a state dict into a target (signature (target, state)); it does not serialize to a file. Saving a checkpoint through this callback raises.

Location: braintools/trainer/_callbacks.py (ModelCheckpoint._save_checkpoint).

Fix: use braintools.file.msgpack_save(filepath, checkpoint).

Found while adding test coverage (#90).

[chaoming0625]

Fixed in #98 (merged to main, commit b8ae6bd; CI green). The PR's "Closes #92, #93, …" only auto-closed #92 because GitHub requires a closing keyword before each issue number — closing manually.