Handle non-contiguous iterators in reproducible reductions

[gevtushenko]

As we kept simplifying the reproducible reduction kernel, we removed the code path responsible for handling non-contiguous iterators. Reproducible reduction should correctly handle fancy iterators.

This issue can be closed by:

1. Passing test case of reproducible reduction on thrust::transform_iterator or thrust::constant_iterator
2. NVBench result illustrating that this change doesn't regress performance