cuda.compute: implement functionality needed for Awkward (jagged) arrays

This meta issue tracks work related to supporting operations on Awkward arrays.

A starting point for this effort is the initial example in https://github.com/scikit-hep/awkward/pull/3734. While this functionality can be implemented in terms of existing `cuda.compute` primitives at the time of writing, we can optimize even further by exposing and implementing more CUB/Thrust functionality.

# Algorithms needed

## P0:

These can't really be expressed efficiently with existing algorithms exposed in `cuda.compute`:

- [x] #6695
- [x] #6688

## P1:

These can be expressed somewhat efficiently; maybe requiring more passes on the data than necessary:

- [ ] #7123
- [ ] #7125

## P2:

These are nice to haves:

- [ ] #7124
- [ ] #7126
- [ ] #7127