Feature request: Add MCP Sampling support

Hi team, MCP sampling is the protocol extension that lets an MCP server ask the client to perform an LLM completion and return the result back to the server. In practice, that means: The server sends a sampling_request message containing a prompt and optional parameters; The client decides which model to call, enforces safety/quotas, runs the completion, and replies with sampling_response. Here is the documentation from MCP: https://modelcontextprotocol.io/specification/2025-03-26/client/sampling Happy to draft a proof-of-concept or open a PR if that helps move things forward. Let me know what you think!