Unify accumulator type handling across CUB/Thrust

[bernhardmgruber]

I vaguely remember that not every reduce or scan algorithm uses ::cuda::std::__accumulator_t to determine the accumulator type to use. We should consolidate this behavior.

[bernhardmgruber]

@gevtushenko said there is divergence between CUB and Thrust. CUB uses __accumulator_t, whereas Thrust uses the type of the initial value.

[bernhardmgruber]

I had another look and realized that the C++ standard seems to determine the accumulator type to either be the iterator value type or the initial value type. So it seems the divergence between CUB and Thrust is fine.

[fbusato]

This is also related and relevant to SIMD reduction. Using cuda::std::plus<> vs. cuda::std::plus<T> could affect performance. e.g. cuda::std::plus<> applied to int16_t induces implicit promotion which disables SIMD