perf: give a major rework on how our deleting works

[Davidson-Souza]

This pr improves accumulator deletion significantly in a few ways.

Our deleting code had some major bottlenecks, like calling calculate
hashes twice (one for the old roots and one for updated ones) and
filtering the nodes that are in the proof, which was causing a linear
search on proof_positins for each element in nodes.

This commit fixes that by:

• Creating a calculate_nodes_delete function that computes in parallel
  both updated and current roots, so we don't need to loop twice. In
  the future, we could even use vectorized computation of sha512_256
  two further speed things up.
• Don't filter proof positions from the nodes vector in
  calculate_nodes_delete. This will waste a little memory as nodes in
  the proofs are returned, but won't cause a new allocation as we are
  returning nodes as is.
• calculate_hashes now won't delete, so it gets simplified to only
  work when checking or updating proofs.

After running the new code for two days with floresta, the speedup is
clear, with calculate_nodes_delete taking less than 2% of the CPU time
for block validation. Before this path, it was using >40%. Here's a before and after