fix(cli): close stateless Happy MCP server on response close

[liunice]

Summary

Fix the built-in Happy MCP bridge for stateless HTTP mode by aligning its lifecycle with the MCP SDK example:

• create a fresh StreamableHTTPServerTransport per request
• create a fresh McpServer per request
• close both on res.close instead of reusing them across requests

Problem

On Linux/NAS environments, Happy's built-in mcp__happy__change_title tool could fail with errors like:

• No matching deferred tools found
• No such tool available: mcp__happy__change_title

The root cause is that the stateless HTTP MCP implementation in startHappyServer was reusing a single StreamableHTTPServerTransport (and connected MCP server state) across multiple HTTP requests.

Why this fix

The MCP SDK's stateless Streamable HTTP example creates a new server/transport for each request and performs cleanup on response close. This PR makes Happy follow the same lifecycle pattern.

Validation

Tested on a Synology NAS / Linux environment where the bug was reproducible:

• before: mcp__happy__change_title failed reliably
• after: title updates succeed again
• takeover UI still works as expected after remote control is attached

Also discussed in #564.