Fix A_mul_B! for y inputs with NaNs by setting elements to zero when β is zero.

[andreasnoack]

Fix JuliaLang/LinearAlgebra.jl#93

[jiahao]

This looks fine. Possibly irrelevant,but I'm still concerned at the inconsistent interface of A_mul_B! (#5332). Sometimes it's a synonym for _axpy, i.e. it computes A*x+y and stores it in y, like in these particular methods, and at other times it's literally A times B, overwriting B.

[andreasnoack]

Updated to use scale! and fill!.

First of all, the different interfaces for A_mul_B! are also for different types, so I am not deeply concerned. When using a mutating function, the general advice must be to read the documentation to figure out what is getting mutated. Some of the differences are also well motivated, e.g. that applying a rotation works directly on the factor in the multiplication. However, we should strive for consistency when possible. The difference between A_mul_B!(Strided,Strided,Strided) and A_mul_B!(Number,Sparse,Strided,Number,Strided) is more troubling to me. Actually so much that I've tried to change it.