Dedicated ReductionEngine for raw arrays

[Algomorph]

This can be used to address multiple issues in the code in the future, e.g. valid depth counting in image processing and the like.

The engine should accept two lambdas:

1. Extract initial count from a single raw array element
2. Combine two previous results together to yeild a third

The warpReduce algorithm from the original InfiniTAM is useful to count up shared CUDA values.

The CPU implementation should be implemented similar to the VolumeReduction, and also take into account the immense CPU caching capabilities when traversing a single contiguous array.