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

[Undiscussed Horrific Abuse, Victim & Survivor of]

Meanwhile, in the transformers library, I've compartmentalised calls
to Amin Rezaei's code into the modeling_utils source file, so I can do
most of the work whether or not the pull request is accepted. If it's
not replied to in a timely manner, I can copy the MIT-licenseed
implementation (which is a direct copy from the paper) into
huggingface's repository later.