Add a GPU calculation backend

[hav-gerdint]

Cumulative impact calculations are highly suitable for parallell processing since calculations in a pixel is independent of adjacent pixels (which is why the current implementation is a PointOp in the JAI API).

The best calculation backend will thus be GPU-based since the massive amounts of cores could be expected to be utilized very well. Estimated performance increase is several orders of magnitude, and probably become largely IO-bound. Ideally all raster data would fit in GPU memory, which should enable extremely fast calculations.

OpenCL or some API that would abstract away different GPU vendors differences would be best, but straight CUDA would also be very good (and perhaps fastest), and most server-based GPUs tend to come from Nvidia at the time of writing.

A CUDA-based backend was developed has been developed in another Symphony implementation which has shown very good performance characteristics.