Bounds error when tiling

[athas]

We may generate out-of-bounds reads when doing 2D tiling for inputs that are smaller in the outer dimension than the tile size. E.g. a standard matrix multiplication will have problems when given inputs of size [31][32] by [32][32].

[athas]

There is a straightforward fix for this. Unfortunately, it will reduce the performance of 2D block tiled kernels, like matrix multiplication. I think we just have to accept that for now, and see if we can recover the performance later.

[athas]

The slowdown is about 2%.