fix: save and validate at final FSDP step boundary

[number-eleven11]

Summary

Run the existing checkpoint and validation interval predicates once more when train_loop reaches num_training_steps. This covers a final optimizer step that lands exactly on an interval boundary.

Fixes #124.

Why

Inside the loop, the save and validation checks run before global_step += 1. On the final iteration of a 100-step run, those checks still see step 99. The counter then becomes 100 and fit() exits before its epoch-end save and validation.

The added check runs only when the configured training limit has been reached, so normal intermediate interval handling is unchanged.

Reproduction / verification

• Configure num_training_steps: 100, save_interval: 100, val_interval: 100, and ignore_until_interval: 0 in a working FSDP training setup.
• Before the fix, the run exits without the step-100 checkpoint. With the fix, a 100-step training smoke run completed and wrote 0_100/model.safetensors; final interval validation also ran.
• python3 -m py_compile wall_x/trainer/fsdp_trainer/fsdp_trainer.py
• python3 -m pytest tests -q (1 passed, 1 skipped)
• git diff --check

Only wall_x/trainer/fsdp_trainer/fsdp_trainer.py is changed.