Avoid duplicate final checkpoint save

[morluto]

What

BaseTrainer.train() saves a checkpoint at every configured checkpoint interval, then unconditionally saves again after the training loop.

When the final training step falls on a checkpoint interval, this writes the same step_<N> checkpoint twice and submits automatic evaluation twice.

Change

Only perform the post-loop save when the final step was not already checkpointed.

Runs that finish between checkpoint intervals still save their final checkpoint as before.