to make it consistent with the other options

In the benchmarks, this gives a quite good speed increase, but in
real-life workloads, the speedup is not as stark (from ~66s to ~61s
here).

What is maybe a neat secondary effect here is that we don't end up with
empty tiles, as we only allocate (in tile_mut) when there is a nonzero
pixel to blit. This decreases the number of saved tiles, especially on
larger zooms.