Simplify replicate_model_fn.GatheringOptimizer inspired by the past comments.

- I worked around the need to rely on Optimizer.__class__ for keeping track of the gradients. Now we are relying on the order of collecting them.  I also added a basic verification that all towers have indeed called the same number of optimizers.
- I now allow the user to increment global step however many times they wish.

The changes above allowed me to support using the same optimizer class multiple times in a tower.

I also renamed GatheringOptimizer to TowerOptimizer, which is a breaking change.  #lifeincontrib

PiperOrigin-RevId: 181381569