Replace critnib with a RAVL tree to track allocations in providers

[lplewa]

This is a rough idea - I'm not 100% sure that this solution will not have any disadvantages.

Currently, on every split or merge, we add and remove entries from critnib structures to track allocations at various levels. This proposal suggests replacing those structures with a RAVL tree instead.

A key advantage of using a RAVL tree is the ability to search for a key that is equal to or lower than a specified value. After a split, this feature allows reusing the original entry in the tree, eliminating the need to update it. As a result, overall performance is improved by reducing the overhead of frequent insertions and deletions. In addition, a RAVL tree will be smaller, which further enhances search times.

[igchor]

I believe that currently, we do update + insert in the split and only update in merge (see https://github.com/oneapi-src/unified-memory-framework/blob/main/src/provider/provider_tracking.c#L318C30-L318C66). How would using RAVL improve this?

Critnib also allows you to search for key that is equal or lower: https://github.com/oneapi-src/unified-memory-framework/blob/main/src/critnib/critnib.h#L36

Also, I'm not sure if RAVL would be smaller - it's a binary tree and critnib has a branching factor of 16 (although the number of level for critnib will depend on how similar are the keys).