[FEA] Expose make_monotonic in pylibraft

cuML users would like estimators to accept non-consecutive target values outside of 0, n), which can occur in scenarios such as training on imbalanced datasets with shuffled data and cross-validation.

cuML has a make_monotonic utility to do this based on CuPy rawkernels, but this requires JIT overhead. We could instead use the make_monotonic primitive in RAFT, which was implemented for DBSCAN in cuML but moved to RAFT.

We can expose this primitive in pylibraft and use it in cuML to both avoid the performance overhead and clean up the cuML codebase.

In case it's relevant, there is also a separate request for renumbering for a possibly different use case.

Start by locating the existing RAFT make_monotonic primitive and the pylibraft binding entry points. Review how similar RAFT primitives are exposed and tested; done means cuML can access make_monotonic through pylibraft without the CuPy rawkernel JIT overhead.