feat: refactor ai agent executor

[pikann]

Summary

• services/ai-agent/src/agent/executor.py — refactored the conversation-run/event-persistence pipeline after reading the installed openhands-sdk 1.30.0 source directly:
  • Removed the dead _make_token_callback / _TurnState machinery. RemoteConversation.__init__'s real constructor has no token_callbacks parameter — it's silently absorbed by a **_: object catch-all (only the Conversation factory's type-checking overload claims to support it), so this code had never fired in production; every agent reply was always persisted through _persist_event's fallback path anyway. Zero behavior change, ~90 fewer lines.
  • Throttled the events.reconcile() workaround (added to unblock the WS-client-thread-dies-permanently upstream bug, OpenHands/software-agent-sdk#1532) from every 2s poll to every 10s. reconcile() re-walks the entire event history over REST with no incremental cursor, and the SDK's own internal blocking run() only pays that cost once, at completion — not on every tick. The cheap execution_status poll stays at 2s for stop-responsiveness; a final reconcile() is still guaranteed on every exit path (done/error/timeout/manual-stop) so a run can never end with an un-persisted tail.
  • Swapped the hand-maintained "done" status set for the SDK's own ConversationExecutionStatus.is_terminal().
• services/ai-agent/src/agent/repo_tools.py — removed the built-in create_pull_request tool. PR/MR creation, review, and commenting is repository-plugin-specific and is handled entirely by each plugin's own tools now, not a built-in agent tool.
• scripts/install-local-plugin.sh — added support for building and installing a plugin's MCP bundle (mcp/dist/mcp.js) alongside the existing backend/frontend artifacts, for plugins that define an mcp/ directory.

Type of Change

• Other (refactor / internal cleanup)

Test plan

• pytest tests/ --ignore=tests/e2e — 83 passed
• ruff check — clean
• PACA_E2E=1 pytest tests/e2e — requires Docker; not run in this environment, recommend running before merge since it exercises the real sandbox + WebSocket path this PR touches

Checklist

• The change is focused and scoped per area above.
• No behavior change intended in the executor beyond the reconcile-interval widening (2s → 10s), called out explicitly above.