do GLU gating for attention layer output, without queries

create a cross attention only attention layer (CrossAttender)

allow for only cross attention in attention layers

allow encoder to cross attend

default to topk sampling

bump version

fix talking heads

Update README.md

Update README.md

cleanup