Added more providers and provider refactor

[zfoong]

What and why:
Currently, we are providing only a limited number of LLM providers, and with how the interface was designed, it is not easy to add another one. Also, KV caching is breaking due to the trigger modifying the prompt above the event stream block, causing mass cache misses.

Item and features added:

• Point Moonshot and Minimax provider to international endpoint.
• Fixed cache misses caused by trigger
• Added new 12 providers: Groq, Mistral, Together AI, Fireworks, Cerebras, Qwen (Alibaba), Hugging Face, NVIDIA NIMPerplexity, LM Studio (local), vLLM (local), llama.cpp (local).
• Updated AGENT.md and improved model switching using chat.

[zfoong]

Cost can be inferred by token usage.

[zfoong]

Session title can change after the session title is decided AFTER the first request, causing cache miss of the first request.