Support unaligned workloads in reproducible reduction

[gevtushenko]

As we kept simplifying the reproducible reduction kernel, we removed the code path that handles contiguous iterators that are not aligned at 16 bytes (float4). We should investigate the following options:

1. Return runtime check of input pointer alignment: if this change doesn't introduce performance regressions, we should select it (best approach from the compilation time perspective).
2. Split reproducible reduction kernel into 16-bytes aligned and 4-bytes aligned versions: this option should be selected only if the first option leads to performance regressions of more than 5%. This approach will lead to kernel duplication and slowest compilation times.
3. Get rid of the vectorized input processing

This issue can be closed by:

• a passed test case of reproducible reduction with thrust::device_vector<float> vec(n); reduce(vec.begin() + 1, n - 1)
• NVBench result illustrating absence of performance regression from this change

[SAtacker]

We do not need aligned iterators for the simple kernel as far as I know. Should we close this?