fix: serialize ACP requests to prevent session/prompt race condition (#499)

[MrRealORG]

Summary

Fixes a race condition in mimo acp mode where session/prompt fails with "Session not found" even when using a valid sessionId just returned by session/new.

Root Cause

The @agentclientprotocol/sdk calls #processMessage() without await in its #receive() loop. This means rapid JSON-RPC requests are processed concurrently, not sequentially. When session/new yields at an async HTTP call (sdk.session.create()), the event loop is free to process the next message (session/prompt), which looks up the session in the in-memory Map — but the session hasn't been stored yet because the HTTP call hasn't completed.

t0: session/new  → starts async HTTP call (yields to event loop)
t1: session/prompt → looks up session in Map → NOT FOUND ❌
t2: session/new HTTP response arrives → session stored in Map ✓ (too late)

Fix

Add a lightweight promise-chain mutex (serialize()) to the Agent class that ensures all session-affecting ACP requests are processed sequentially. This is a minimal, non-invasive workaround for the SDK's fire-and-forget behavior.

Methods wrapped:

• newSession / _newSession
• loadSession / _loadSession
• prompt / _prompt
• cancel / _cancel
• setSessionMode / _setSessionMode
• setSessionConfigOption / _setSessionConfigOption
• unstable_setSessionModel / _unstable_setSessionModel

Testing

The fix can be verified by sending rapid ACP requests via stdin:

echo -e '{"jsonrpc":"2.0","id":0,"method":"initialize","params":{...}}\n{"jsonrpc":"2.0","id":1,"method":"session/new","params":{...}}\n{"jsonrpc":"2.0","id":2,"method":"session/prompt","params":{"sessionId":"<from above>","prompt":[...]}}' | mimo acp

Fixes #499