Targeted stable_muladdmul branches in small matmul

[jishnub]

This PR changes the signature of matmul2x2! and matmul3x3! so that they now accept the alpha and beta parameters instead of the MulAddMul. Consequently, the @stable_muladdmul macro may be applied to smaller blocks of code within the functions instead of applying it to the entire call. This reduces the matrix multiplication latency by a bit.

julia> using LinearAlgebra

julia> A = rand(2,2); B = rand(size(A)...); C = zeros(size(A));

julia> @time mul!(C, A, B);
  0.903632 seconds (2.54 M allocations: 130.322 MiB, 5.52% gc time, 100.00% compilation time) # master
  0.813335 seconds (2.28 M allocations: 117.025 MiB, 2.07% gc time, 100.00% compilation time) # This PR

[dkarrasch]

Awesome!