FlexLLM server demo

[goliaro]

Description of changes:

TODOs:

• Streamlit app
• Chat protocol
• Add LLAMA-3.1 and Llama3.2 support & check alignment (in particular, RoPe)
• Support LoRA in attention projections (Attention projections (QKV, O) disaggregation #1436)
• Be able to add LoRA layers at runtime, and deallocate memory when done
• Be able to set parameters (e.g. max sequence length) at runtime for each request
• Be able to set generation configs (top_p, temperature, etc) at runtime for each request

Related Issues:

Linked Issues:

• Issue #

Issues closed by this PR:

• Closes #

---

This change is