Don't zero out destination in generic_matvecmul

[jishnub]

Currently, mul!(c::AbstractVector, A::AbstractMatrix, b::AbstractVector) is computed (schematically) as

c .= 0 # zero of the correct type
for y in eachcol(A .* permutedims(b))
    c .+= y
end

Filling c with zeros at the beginning is unnecessary. We may instead start by storing first(eachcol(A .* permutedims(b))) directly in c, followed by a loop over the other columns. This reduces one loop over c, although the difference in performance is minimal.

[dkarrasch]

I don't really see how to get this working, without modifying the corresponding test case, which fails for all commits on this branch. We need this sum here to be able to promote to the eltype of C, but we don't have zero available for neither A nor B. If this doesn't affect performance too much, maybe it's not worth it after all?

[jishnub]

Yes, probably not worth it