Mempool optimizations

[ImplOfAnImpl]

• Collections in MempoolStore that have tx id as the key were changed from btree-based to hash-based. Note that I decided to use hashbrown's containers directly, instead of using the standard ones (which also use hashbrown under the hood), because the size estimation is more precise in this case.
• There were a few places where the whole mempool would be traversed when adding a new tx:
  • TxPool::remove_expired_transactions would iterate the entire MempoolStore::txs_by_creation_time even though only the first few txs need to be checked.
  • MempoolStore had methods refresh_ancestors/refresh_descendants, which were called from add_to_descendant_score_index/remove_from_ancestor_score_index/remove_from_descendant_score_index - they would remove all ancestors/descendants from txs_by_ancestor_score or txs_by_descendant_score and reinsert them with the new score. I removed these refresh_ methods, and now txs_by_ancestor_score/txs_by_descendant_score are updated in the places where the score is modified - in update_ancestor_state_for_add/update_ancestor_state_for_drop/update_descendant_state_for_drop.
• New tx's ancestors were collected multiple times. I added the TxMempoolEntryWithAncestors struct, so that the collected ancestors can be re-used at least during the new entry initialization.

I also updated comments and did some renaming in get_best_tx_ids_by_score_and_ancestry and its helper function collect_selected_ancestors (which is now called collect_nearest_selected_ancestors and is in a separate module). The logic here wasn't changed, it's just that the old comments and naming were misleading.