Incorporate Mia's codebase

[zkokaja]

Major changes:

• map any model to any tokenizer
• use automap for device usage
• largest bloom model needs downloading
• perplexity calculation

[zkokaja]

Do we need to align a model with the 20b version with LCS every time? Or can we can use the new in_model column in the pickles.

[zkokaja]

Do this in combination with #122