Support prompt caching for claude models

[liangjs]

Do we plan to support the prompt caching feature for the claude models?
This can reduce the cost when the agent runs many long steps.

It requires to set a special cache_control in the chat messages.
Referece: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching

[cangozpi]

I’ve submitted this PR to introduce caching support for Gemini models. If approved, I plan to follow up with support for Claude models too.