support training, fix some inplace func of nn

[Gy-Lu]

This PR should be merged after the other two.

1. Add training code and its shell script, support DDP now.
2. Add lr_scheduler and move loss to hub.
3. Modify some bugs introduced by openfold, the inplace func in dropout and the unused out_product_mean in evoformer.

For unit test, it would cause OOM on CI machine. I am trying to develop a tiny model.

[Shenggan]

Seem still have some bugs:

1. from fastfold.model.loss import * -> from fastfold.model.hub.loss import * in fastfold/model/nn/heads.py

2. FastFold/fastfold/model/fastnn/template.py

   Line 394 in 164f677

   t = self.layer_norm(t[i])

   introduced by refactor chunk for training #117