Fix mamba integration by making it a variant of outlines.models.transformers

[lapp0]

Fixes dottxt-ai#808

Rendered docs: https://github.com/lapp0/outlines/blob/fix-mamba-integration/docs/reference/models/transformers.md#alternative-model-classes

Problem

• Per the above issue, Mamba doesn't use logits processors, it uses SequenceGenerator. It should use SequenceGeneratorAdapter and have logits processors manage the automata.
• models.mamba doesn't work at all in main.

Solution

Update models.transformers to accept a model_class argument allowing for additional model types beyond AutoModelForCausalLM. Make models.mamba simply a variant of models.transformers with the model_class=transformers.MambaForCausalLM passed. This results in nearly zero maintenance required for models.mamba.

Unrelated work

Additionally Zach on Discord requested T5-based structured generation. I tested it with model_class=transformers.AutoModelForSeq2SeqLM and it works with zero additional changes. The only change I made related to this is adding a model_t5 fixture to our test_generate.py tested models and documenting it in docs/reference/models/transformers.md.