[EPIC] Utilize nondeterministic reduce further where possible

The new nondeterministic reduce can currently only be used by specifying the `not_guaranteed_t` determinism requirement in the new env based `DeviceReduce` API. Additionally, it only supports `cuda::std::plus<>` currently. We need to take more advantage of the new reduction implementation where possible:

- [ ] Default to `not_guaranteed_t` for integer reduction since integer reduction is always deterministic
- [ ] Add support for other operations, such as min, max, mul, etc.
- [ ] Expose nondeterministic reduction to c.parallel
- [ ] Expose nondeterministic reduction to cuda.cccl.parallel