Improve inferrability of generic matrix-vector multiplication

[chethega]

This hoists some work-arounds for computation of eltypes for zero-size vectors out of the loops in order to get a speedup in some cases where type inference otherwise takes tuple types.

Thanks to #36941 for opening the issue.

Before:

julia> using BenchmarkTools
julia> N = 100;
julia> x0 = ones(N);
julia> A_t = transpose(ones(Int64, N, 5*N));
julia> @btime $A_t * $x0;
  103.525 μs (1 allocation: 4.06 KiB)

After:

julia> @btime $A_t * $x0;
  34.917 μs (1 allocation: 4.06 KiB)

I put partial blame on #35164. I verified that #35163 stays fixed.