Instead of locking for the whole time the undo data is being stored,
only lock while the step's per-node undo node map is being accessed.
This is fine because each PBVH node is only processed by a single thread.
Changing the node vector to not store anything until the undo step is
finalized makes this process a bit simpler because we don't have to build
both the map and the vector at the same time.
Overall this improved the performance of the sculpt brush benchmark
from #118145 by 12%, from 0.68s to 0.61s.
889b142924