Mempool: fix spuriouly failing test; shrink hash tables; some cleanup.

[ImplOfAnImpl]

• Since the switch to using hashbrown tables in mempool, the test mempool_full_real can fail spuriously, because now removing a tx from a full mempool and inserting one with the same size is not guaranteed to succeed (this is because hashbrown tables leave a tombstone instead of a removed element, which may not always be re-used during insertion). So the test had to be relaxed.

• Mempool will now try to shrink its hashtables' capacities from time to time.
  Note that I initially implemented this to handle a degenerate case where an attacker would fill the mempool with 200'000 small transactions, which may permanently occupy up to 40Mb worth of capacity in the hashtables, thus reducing the effective mempool size by the same 40Mb. But then I noticed that we actually recreate the whole mempool store each time the mempool reorgs to a new tip, so the degenerate case may exist for several minutes at best. But IMO it's better to keep the shrinkage logic, at least as a fail-safe.

• Small cleanup - MempoolStore's fields are now non-public.