SpGEMM: Fixed problem in CUDA kernel (stage 3) due to CUDA bug(?)

Same problem showed up with OpenCL earlier in
216a6ac4
I assume that we are hitting a bug in the CUDA stack here,
since the problem only shows up on some CUDA devices (e.g. K20m)
and only with certain build configurations. A debug build, for example,
does not show any issues.

See also the follow-up discussion in #147.