Feature request - save every n steps

For big datasets an epoch can take a lot of hours. It would be much more practical if checkpoints can be saved every n steps. After looking at the code I think the modification should be rather straightforward.

Start by tracing the existing checkpoint-saving logic and the training entry points that count epochs and steps. Confirm how an n-step interval should interact with current epoch-based saves, then verify that checkpoints are written at the requested interval without breaking existing behavior.

Written by the indexing model from the issue text.