Improve cache-friendliness of cartesian products

[clbarnes]

1. Sort query and target indices (possibly align them?)
2. Address in Morton (or Hilbert) order

• will this actually help? Or does the fact that we're looking up into a vec break this anyway?

[aschampion]

I recalled testing a dirty version of this and found an old branch here. Unless there's a log of a slack thread or something the only note about the outcome is from #17:

Z-order scheduling of the neuron matrix parallelization for cache obliviousness did not help in FIB250, but the total data size is quite small relative to cache size. @clbarnes will produce a larger dataset from the larval whole brain at some point.