Fix threadblocks in Errf_CUDA

[miromarszal]

Currently threads are divided into N block of N threads per block, where N is the width of an array. This has a very low limit on scalability, i.e. the maximum number of threads per block.

Rewrite in such a way, that the code scales efficiently at least for arrays whose size is a power of 2. Perhaps dividing into square submatrices would be still quite readable, but much more scalable.