update requirements for ReducerArgumentNonConst

[nmm0]

Parallel reduce requirements say:

If ExecPolicy is TeamThreadRange a “reducing” functor is not allowed and the ReducerArgument must satisfy the Reducer concept or ReducerArgumentNonConst must be a POD type with operator += and operator = or a Kokkos::View

Rather we would like requirements that reflect that the type is:

• default constructible
• copyable
• and "behaves ok when cudaMemcpy is called on it"

[nmm0]

sycl has an applicable requirement "Device copyable": https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#sec::device.copyable

[ajpowelsnl]

TODO:

• Check parallel_reduce documentation, and verify behavior that is accurately described (i.e., POD types are no longer a thing in a thing);
• Remove mention of POD (for C++ 20 and beyond) info from our documentation; a POD object does not have user-defined constructors
• Understand if current behavior (above) is the same as desired behavior
• Please verify that if a CPU pointer is copied to device, you get "correct" behavior; i.e., the type (copied from CPU) must be device-usable and is correct, without pointers to host memory spaces