Handle long contexts better

[pseudotensor]

While had added critical protection against failure when inputs too long, still if prompt happens to be too long, tokenization will do truncation=True for a max_length in some way (left or right or whatever), and can drop prompting of : etc.

Right now form prompt, then tokenize. But need to tokenize first ourselves (even for h2oai_pipeline), and remove non-human non-bot parts using split or similar, remove some stuff to get token length down, and then join back up.

Else missing human truncated out confuses model that just sees : and will repeat answers etc with another :. The output is cleaned from that, but stopping doesn't stop from that so calculations go in background and lag user.

[pseudotensor]

#255 helps for h2oai_pipeline.py use. Should solve 99% of issues, but still fudge w.r.t. prompt itself.