Completely workover the m_chainsMutex locking, moving the locks around so all the expensive I/O operations happen without the central mutex locked.

This significantly reduces the amount of UI lockups caused by background-parsing or missing-include completion.