This PR is a more advanced form of the dirty-to-clean cacher (#18995), where we reuse previous database write batches as datasets to uncache, saving a dirty-trie-iteration and a dirty-trie-rlp-reencoding per block.