fix: close Browserbase sessions on HTTP disconnect

[RitwijParmar]

What changed

Closes #187.

This wires the per-client HTTP MCP lifecycle back to the Browserbase session lifecycle:

• registers each MCP server's SessionManager.closeAllSessions() with ServerList
• makes transport.onclose call serverList.close(server) instead of only removing the transport routing entry
• makes ServerList.close() idempotent/re-entrant safe, so server.close() triggering transport.onclose does not double-close the same server/session
• updates closeAll() to use the same cleanup path as single-server shutdown

Why

In long-running --port mode, a client can disconnect without calling the end tool. Before this change, the transport entry was removed but the MCP server, SessionManager, Stagehand instance, and Browserbase cloud session could stay alive until timeout; with keepAlive, that can turn into a real session-minute leak.

Validation

• ./node_modules/.bin/tsc --noEmit
• ./node_modules/.bin/eslint src/server.ts src/index.ts src/transport.ts src/server.test.ts

I also added focused Vitest coverage for cleanup ordering and re-entrant close behavior. I couldn't run Vitest locally because macOS rejected the downloaded Rolldown native binding in this checkout, but the test is included for CI.