`cub::ThreadStore` reinterprets non-trivially copyable types

When `cub::ThreadStore` is used to store a `T` that is not trivially copyable, the value to be stored is still reinterpret-casted and stored using as many device words as needed. This violates the type requirements on `T`. We should probably add a fallback that performs an ordinary store if `T` is not trivially copyable.

This occurred when working on #6811 for the type `custom_output_t` in test `catch2_test_device_scan_iterators.cu`.