Add use_cache option

[wiseodd]

Huggingface has this option set to True by default.

https://github.com/huggingface/transformers/blob/82486e5995ed0a65520b10ce1ea938214a199231/src/transformers/generation/configuration_utils.py#L131-L133

It caches the previous "key" & "value" attention activations, so during autoregressive text generation, one only needs to freshly compute the "query" part of the attention.