fix local windowed attention to be overlapping

value residual learning (lucidrains#312)

* cite

* add value residual learning

* oops

* slip in value residual learning for pairformer stack

* also cite Nguyen, whose initial paper led here

every machine keep track of ema if using switch ema

a new paper claims there is a free lunch by setting model weights to …

…ema weights every epoch. allow researchers to experiment with this, conveniently already available in EMA-pytorch due to hare and tortoise paper

fix MSA_TYPE