SpGEMM: Reduced memory footprint of CUDA implementation.

Maximum number of elements in scratchpad was too pessimistic.
Current estimate only depends on the maximum number of nonzeros per row in B processed by the block.