Small performance gain in allocations

As measured in
https://gist.github.com/kirushik/e0d93759b0cd102f814408595c20a9d0,
it's much faster not to iterate over zeroes, and just allocate a
contiguous array of zeroes directly.