Optimizer is reset every epoch rather than during training stage switch

In the paper, it is mentioned that "we also reset the optimizer state when training stages switch."

However, in
https://github.com/facebookresearch/coconut/blob/d1c10b134ad2bded92b0eca4c0a422c513a12dea/run.py#L312-L320
the optimizer is reset every epoch.

Start in run.py around lines 312-320 and compare the optimizer reset there with the paper's statement about resetting when training stages switch. Trace where the training-stage transition is handled, then verify that the optimizer is reset at stage changes rather than at every epoch; confirm the behavior with the relevant training run or tests if available.

Written by the indexing model from the issue text.