Currently a single thread dispatches events to all sessions.

This may become a bottleneck, and introduces unnecessary buffering and complexity (e.g., ensuring the recording has been read far enough so that a session can be fed response data to allow statistical comparisons when a replayed operation ends).