Narrower MulAddMul stabilization in bidiag mul!

[jishnub]

This builds on #56360, so it would be easier to review by rebasing after that one is merged.

This PR reduces latency for mul! involving Bidiagonal matrices.

julia> using LinearAlgebra

julia> B = Bidiagonal(1:4, 1:3, :U); D = Diagonal(B); C = similar(B, size(B));

julia> @time mul!(C, B, D, 2, 2);
  0.117434 seconds (243.24 k allocations: 12.758 MiB, 99.97% compilation time) # nightly
  0.054078 seconds (122.71 k allocations: 6.076 MiB, 99.94% compilation time)

julia> @time mul!(C, B, B, 2, 2);
  0.716847 seconds (1.89 M allocations: 93.264 MiB, 16.74% gc time, 99.99% compilation time) # nightly
  0.493675 seconds (1.04 M allocations: 52.477 MiB, 24.34% gc time, 99.99% compilation time) # this PR

julia> B = Bidiagonal(1:4, 1:3, :U); A = Array(B); C = similar(B, size(B));

julia> @time mul!(C, A, B, 2, 2);
  0.287015 seconds (325.09 k allocations: 16.715 MiB, 10.32% gc time, 99.95% compilation time) # nightly
  0.220946 seconds (250.40 k allocations: 12.395 MiB, 15.51% gc time, 99.99% compilation time) # this PR

julia> B = Bidiagonal(1:4, 1:3, :U); v = rand(size(B,2)); C = similar(v);

julia> @time mul!(C, B, v, 2, 2);
  0.351700 seconds (570.67 k allocations: 30.650 MiB, 8.00% gc time, 99.99% compilation time) # nightly
  0.272777 seconds (901.02 k allocations: 47.221 MiB, 10.34% gc time, 99.99% compilation time) # this PR