CUDA: set_rows + cpy.cu refactor

[am17an]

To also support the rest of the quantized data-types in set-rows

Moving cpy functions to a common headers to re-use in set-rows, I will do get-rows next if this refactor seems the right direction. I feel we can still make the interface a little cleaner

[JohannesGaessler]

Long-term we should rewrite the quantization code to make use of coalesced memory accesses. For an example, see quantize_q8_1 in quantize.cu, that kernel loads 32 contiguous floats and writes 32 contiguous bytes for the quantized values.