feat: add prompt_cache_key for cross-call cache routing

[san0808]

Summary

Adds prompt_cache_key=agent_id to all OpenAI and Azure LLM API call sites so that calls for the same agent route to the same GPU server, enabling cross-call cache hits on the system prompt.

Previously, each new call landed on a random server and started with a cold cache on the first turn. Now the first turn reuses the cached system prompt from previous calls for the same agent.

Caller-controlled opt-in: bolna sets prompt_cache_key only when agent_id is explicitly passed in kwargs. See https://github.com/bolna-ai/dashboard-backend/pull/2041 for the allowlist consumer.

What changed

File	Change
task_manager.py	Forward agent_id to LLM when caller passes it in kwargs
openai_llm.py	extra_body with prompt_cache_key in streaming and non-streaming chat completions; pop and merge for WebSocket path
azure_llm.py	Store agent_id; extra_body in streaming and non-streaming chat completions
openai_base.py	extra_body in _build_responses_create_kwargs and _generate_responses
graph_agent.py, knowledgebase_agent.py	Pass agent_id through to LLM initialization

All changes guarded by if self.agent_id, no effect when absent.