WIP sparse v2

Faster algorithm, but unfortunately not deterministic (because additions can be
made in a different order, depending on relative CPU core loads). The result
can therefore slightly change across runs, because changing the order of
additions in an a multi-element sum can change the result due to floating point
properties (this makes example2.mod fail randomly, because that test checks for
strict equality).

NB: The case B⊗B still needs to be rewritten

[skip ci]