Remove redundant `reduce` implementation

[alliepiper]

thrust/system/cuda/detail/reduce.h contains a complete reduction implementation in the __reduce namespace. In 8702bfe89640c9efcf68cf3ad2b89d4d00e96494, the thrust::cuda_cub::reduce entry points were changed to just call cub::DeviceReduce, but the old implementation remains.

The __reduce implementation is still used to implement the algorithms in extrema.h (minmax_element, etc). These should be updated to use cub::DeviceReduce and the __reduce implementation should be removed.