First submission

fix attention issues for keys != queries

fix pooling issues
now all three types of unet layer have .maxpool
up uses maxpool to get key_padding_mask
rename _get_next_mask -> _get_shrinked_mask

add wandb and data-bin folders to .gigignore

x16 speed up on GPU

Removed separable convolutions
Fixed other minor issues

Debugged

UNet is too slow, but it works