[MCP caching]: pool MCP backends in thread manager

[jgershen-oai]

Depends on #15256.

Summary

This is the PR that actually makes MCP backend reuse happen across loaded threads.

• add a ThreadManager-owned pool of shared MCP backends for enabled Stdio MCP servers
• keep StreamableHttp MCPs, including codex_apps, per-thread/local
• force-rebuild the pooled stdio half on refresh, and reacquire it when sandbox state changes
• add stdio-focused app-server lifecycle coverage for unsubscribe, archive, and archive -> unarchive -> resume

Why

The concrete bug here is duplicate local MCP processes when multiple loaded threads use the same stdio MCP configuration. Pooling only stdio backends fixes that without turning codex_apps into a shared HTTP bottleneck or introducing cross-thread elicitation routing.

Testing

• cargo check -p codex-core
• cargo test -p codex-core mcp_connection_manager::tests::
• cargo test -p codex-core pooled_new_turn_keeps_local_mcp_startup_token_when_sandbox_state_changes
• cargo test -p codex-app-server thread_mcp_pool
• cargo test -p codex-app-server mcp_server_elicitation