This patch extracts fields from `RoomEventCacheState` and move them
into `RoomEventCacheStateLock`. This lock provides 2 methods: `read`
and `write`, respectively to acquire a read-only lock, and a write-only
lock, represented by the `RoomEventCacheStateLockReadGuard` and the
`RoomEventCacheStateLockWriteGuard` types.
All “public” methods on `RoomEventCacheState` now are facade to the read
and write guards.
This refactoring makes the code to compile with the last change in
`EventCacheStore::lock`, which now returns a `EventCacheStoreLockState`.
The next step is to re-load `RoomEventCacheStateLock` when the lock is
dirty! But before doing that, we need this new mechanism to centralise
the management of the store lock.
Instead of creating a fresh client every single time, this creates a
single client with a single event cache store, that's cleared between
runs of the benchmark.
So the tradeoff is:
- we don't have to create a new client anymore, which means no more
async setup code, which means we can avoid using `iter_custom`; a
benefit of this is that this can be benchmarked on CI now.
- but we're also measuring the time it takes to clear the database,
which isn't trivial in itself.
Instead of having one static method duplicating an underlying
`ClientBuilder` method, we can pass the builder directly to a closure,
that will replace it. Call sites are a bit more verbose, but that would
avoid having to add duplicate `MockClientBuilder` methods for each
`ClientBuilder` method.
