BUGFIX: genmatmul! for empty input matrices

[KlausC]

fix JuliaLang/LinearAlgebra.jl#682

Note: This regression was introduced in #33530 .

[andreasnoack]

Why don't we just call _rmul_or_fill!(C, _add.beta) unconditionally. I believe that is essentially the BLAS behavior which ensures proper initialization of C.

[KlausC]

@andreasnoack, I think, because the initialization is only required in the return case.
The rest of the code (when the multiplication actually happens) initializes C by itself. Actually I did not think to change that, because I was focused on removing the bug. Maybe it is worth a separate PR to optimize the code.

[andreasnoack]

I see. In BLAS the β part is handled separately before the multiplication but I now see how it's handled differently here. I think this fix is good to go since the test failures are unrelated.