Splits the state_group_cache in two.

One half contains normal state events; the other contains member events.

The idea is that the lazyloading common case of: "I want a subset of member events plus all of the other state" can be accomplished efficiently by splitting the cache into two, and asking for "all events" from the non-members cache, and "just these keys" from the members cache.  This means we can avoid having to make DictionaryCache aware of these sort of complicated queries, whilst letting LL requests benefit from the caching.

Previously we were unable to sensibly use the caching and had to pull all state from the DB irrespective of the filtering, which made things slow.  Hopefully fixes https://github.com/matrix-org/synapse/issues/3720.

This reverts commit ec716a35.