Handle wide matrices in 2D kernels gracefully.

[benson31]

CUDA has a maximal .y and .z grid dimension of 65536. We should handle cases where the local matrix is wider than that.

[timmoon10]

And if we want to be extra careful, the max .x dimension is 2147483647.

[benson31]

Now that I think about it, if we really want to be technically precise and extremely careful, this is a property of the device in question and its compute capability, IIRC. So really, we should query this in the initialization and use the dynamically determined maximum value in order to achieve the best portability.