make sure masks work with funnel transformer

make sure each bottleneck section in funnel transformer share a diffe…

add funnel transformer

fix encoder feedforward glu

fix bug with encoder relative pos emb

add dropouts

add memory tokens, from Memory Transformers paper

add all-attention memory key/values, set with `attn_num_mem_kv = {int}`

add ability to use entmax15 for sparse attention

bump