Support for loading arbitrary huggingface models (llama, mixtral, gpt_bigcode) for speculative training

[JRosenkranz]

Currently, we require an fms variant of a model, this should make it simpler to load models that have variants that fms does not support natively