[ot][spam][crazy][data] transformer model 'attention' improvement

[Undiscussed Horrific Abuse, Victim & Survivor of]

github/xloem/transformers repo, using return_weights branch of
github/xloem/memory-efficient/attention repo

commit e72d4a3536d1799fc40a85d83a7999e8a39563fc (HEAD ->
memory-efficient-attention, xloem/memory-efficient-attention)
Author: xloem <0xloem at gmail.com>
Date:   Sat Jan 29 07:58:00 2022 +0000

    more bugfixes. output differs, makes sense to step through in
parallel again and compare.