Callback hook for on_after_optimizer_step

Nobody has claimed this yet.

## 🚀 Feature

A callback hook for `on_after_optimizer_step`.

### Motivation

There's a callback hook for `on_before_optimizer_step`, but not for `on_after_optimizer_step`.
That would be useful for implementing an [ExponentialMovingAverage](https://github.com/pytorch/vision/blob/b3cdec1fd808b287e4e0d31f989488c335df1812/references/classification/utils.py#L158) callback: I'd like to update the average weights [after](https://github.com/pytorch/vision/blob/b3cdec1fd808b287e4e0d31f989488c335df1812/references/classification/train.py#L52) the optimizer has updated the parameters. Doing this average weight update with `on_train_batch_end` hook will not be accurate, as the model weights may not get updated after every training batch (due to gradient accumulation).

cc @borda @tchaton @rohitgr7 @carmocca @awaelchli @ninginthecloud @daniellepintz

Start by tracing the existing on_before_optimizer_step callback through the optimizer-step training loop. Compare it with the PyTorch Vision references/classification/utils.py and train.py links to understand the intended post-update timing. Done means the new hook runs after optimizer parameter updates, including with gradient accumulation, and supports the stated averaging use case.

Written by the indexing model from the issue text.