trie: implement lazy loading of trie nodes from disk

[qdm12]

Our memory usage is high since we keep all trie nodes in memory.

Instead we should (like Substrate) store all node hash -> node encoding on disk, and lazy load them as we need to access the trie (both read and write). Only the root hash should be kept in memory, and every node should be retrieved from disk as we go deeper in the trie (to insert or retrieve).

We should also implement pruning for nodes inserted/modified in uncle blocks (and make sure the finalized block doesn't contain them).

We can also implement a cache to keep the most frequently used nodes in memory and prevent decoding them again

Related with #2835 and #2836

[dimartiro]

I found that Ethereum is also using a lazy loading strategy with the node keys -> encoded nodes

https://github.com/ethereum/go-ethereum/blob/d76efbb9be403689bc8d677a16af453728a7a513/trie/triedb/hashdb/database.go#L186-L224

[dimartiro]

In substrate, they are using this trie crate for all trie operations, and they create a TrieDB structure backed by a DB with a cache, and they are implementing the lazy loading there

[dimartiro]

This change will help us to improve our memory usage and our node startup time when we are using an snapshot, also improving our debugging time when we are trying to sync to the tip of the chain and we need to use snapshots for that.

[dimartiro]

This issue will be Done when #3805 is completed

[P1sar]

Design Design considered complete