Use thrust::inclusive_scan for 1D cumsum/cumprod

[colesbury]

For large 1D tensors thrust::inclusive_scan is much faster than our current implementation.

For 4 million elements thrust takes about 430µs vs 200ms for the current implementation.

[pavanky]

May be consider using thrust + for loop for scanning along inner most dimension if the current version is really that slow.

[colesbury]

@pavanky, there are cases where the current implementation would be faster for 2D+ tensors so I didn't want to completely switch to that (if the outer dimension is large enough to parallelize over). If the 2D+ case is a bottleneck we may want to revisit this and possibly pull in segmented scan from moderngpu.

[wickedfoo]

This is the general tradeoff that much of cutorch makes for anything that is a reduction of sorts. Many cutorch reduction functions operate under the assumption that there is a large enough batch (you are doing something along dimension D of an N-dimensional tensor) rather than one thing along a 1-d array. The implementation usually just targets the first (batch) case, leaving bad performance for the non-batch case, which usually requires a different specialization to solve the problem via recursive sub-division or multiple kernel launches with scratch space.