Faster online pruning with cached trie inserted/modified merkle values

[qdm12]

The online pruning finds inserted/modified nodes by going through the entire trie and checking each node for the dirty flag.

We can change that to have each trie cache merkle values (database keys) that were inserted/modified, so it should be much faster.

This should be done after #2835

[qdm12]

When inserting/updating a node in the trie, all the path of nodes from the root to that new or updated node is modified. In other words, for a node at depth N in the trie, N nodes are modified. That means for every Put operation, there are N node encoding and up to N blake2b hashing. As a consequence, performance is severely degraded, so tracking updates in the trie is not feasible. The current walk of the trie will be kept instead, especially since it would be needed to implement the 'write cache' for lazy loading (walk the trie and only write updated nodes since the last snapshot).