[Feature] Support CohereForCausalLM architectures

[accupham]

Would it be possible to add Command-R to the supported models?

It has excellent multi-lingual RAG Capabilities. The architecture is nearly identical to LLama2, and it seems like the equivalent VLLM implementation has many similarities as well.

Here is the discussion in VLLM on implementation details for reference: vllm-project/vllm#3433

[merrymercy]

#369