Fix running parallel_reduce with TeamPolicy for large ranges

[masterleinad]

Fixes #4531. Specifically, the lines

const auto nwork = static_cast<typename Policy::index_type>(m_league_size) * m_team_size;

fix the issue (for CUDA and HIP). The other changes to the types of nwork are just for consistency.

[masterleinad]

Retest this please.

[masterleinad]

This seems to work on all backends now. Note that I had to change the argument type of scratch_space for SYCL to std::size_t to make it work. We should discuss (maybe as a follow-up) if we want to make this consistent for all backends and switch to std::size_t or find a different solutiuon.

[masterleinad]

The changes for SYCL are also discussed in #4551.

[janciesko]

LGTM