Consider persisting/restoring the streaming metadata (last evicted) cache on silo shutdown/startup

[oising]

I had a conversation with @benjaminpetit some months ago about this, and I think we both agreed that this would help with the QueueCacheException storm that may happen during a rolling upgrade or otherwise planned restart of a cluster. The issue is that while the cache is reloaded from checkpoint data, the in-memory dictionary of streamId->lastEvictedSequenceToken is lost. This serves as a secondary mechanism to keep an eye on the n-1,n validation for dropped message detection.