Sometimes LV has to produce really wide vectors,
and sometimes they end up being not powers of two.
As it can be seen from the diff, the cost computation
is currently completely non-sensical in those cases.
I don't really know what i'm doing, but does this look better?
Instead of just scalarizing everything, split/factorize the wide vector
into a number of subvectors, each one having a power-of-two elements,
recurse to get the cost of op on this subvector. Also, check how we'd
legalize this subvector, and if the legalized type is scalar,
also account for the scalarization cost.