fix: isolate chat event loops and persist pending input

[fcsguyxcw]

Summary

• give each AgentContext its own event-loop thread so one blocked chat cannot wedge unrelated chats
• persist each UI message as a pending queue item before dispatch, then drain it in memory so the on-disk copy survives until normal loop completion saves the cleared state
• add regressions for cross-chat loop isolation and pre-dispatch message durability

Root cause

Every chat constructed DeferredTask(thread_name="AgentContext"). EventLoopThread is a singleton by name, so all chats ran on one asyncio loop; one synchronous blocker stopped every chat at once. UI messages were only held in memory before dispatch and chat persistence normally happened at message-loop completion, so a wedge could leave no durable record.

This is separate from #1793 / #1794: those address unsafe teardown of shared loop threads; this change removes cross-chat sharing at the AgentContext boundary.

Fixes #1889

Verification

• pytest tests/test_agent_context_event_loop_isolation.py tests/test_message_durability.py tests/test_defer_lifecycle.py -q — 7 passed
• py_compile passed for all changed Python files
• git diff --check passed

Additional neighboring suites could not collect in the existing local environment because optional pathspec and crontab packages are absent.