[commonware-storage/mmr] Improve forgetting behavior in mem::Mmr

[roberto-bayardo]

mem::MMR currently supports forgetting nodes up to the position of the tallest peak, but in the worst case this still requires we retain almost half of all nodes ever added to the Mmr in memory. We should allow forgetting of old nodes to any point in history, which can be supported by extending in the in-memory MMR to always retain the peak hashes, for example using a map that is independent of the flat nodes vector.

As part of this change we will also want the forgetting functionality to provide proper feedback to the user around how proof generation is affected by returning the index of the oldest-provable element after forgetting. While pruning to the tallest peak allows proving of any remaining element following it, pruning to arbitrary points affects element proving in a less-trivial way.