Starcoder2 model

[jlamypoirier]

The mistral model adapted to run Starcoder 2:

• Use layer norm (RMS still available as option)
• Use standard MLP (gated still available as option)
• Add back biases (optional)
• Change (default?) tokenizer class

Missing for starcoder 1 (do we want to support it?):

• Absolute position embeddings

Other notes:

• Has less entries in modeling auto than gpt bigcode (3 instead of 6), probably doesn't matter
• Using repeat for kv cache in flash attn, might not be necessary.

Still got a bunch of minor things to do (see todos)