Fix syncbn training state

[Marchematics]

Motivation

Preserve the source module's training mode when converting BatchNorm to SyncBatchNorm. Without this, evaluation-mode layers are silently switched to training mode, which can cause incorrect fine-tuning behavior and NCCL timeouts. Fixes #1624.

Modification

Copy module.training to the converted module and add a regression test for evaluation-mode conversion.

BC-breaking

No.

Checklist

1. Targeted lint checks pass.
2. Added test_convert_sync_batchnorm_keeps_training_state.
3. No downstream changes are required.
4. No documentation changes are needed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Marchematics]

Removed the duplicate isfile mock that caused the lint failure in the checkpoint test. The changed test now passes the repository lint checks locally.

[Marchematics]

Targeted regression test passes: tests/test_model/test_convert_sync_batchnorm_training.py.