Suboptimal matrix vector product

There appears to be a scaling problem with matrix vector products, I think to do with vectorization (from inspecting generated code).

Example:

julia> let N=32, M=8
           @btime mul!($(zeros(Float32, M)), $(rand(Float32, M, N)), $(rand(Float32, N)))
           @btime $(rand(SMatrix{M, N, Float32})) * $(rand(SVector{N, Float32}))
           @btime $(rand(SMatrix{M, N, Float32})) * $(rand(SMatrix{N, 1, Float32}))
       end;
  33.646 ns (0 allocations: 0 bytes)
  8.601 ns (0 allocations: 0 bytes)
  9.834 ns (0 allocations: 0 bytes)

julia> let N=64, M=8
           @btime mul!($(zeros(Float32, M)), $(rand(Float32, M, N)), $(rand(Float32, N)))
           @btime $(rand(SMatrix{M, N, Float32})) * $(rand(SVector{N, Float32}))
           @btime $(rand(SMatrix{M, N, Float32})) * $(rand(SMatrix{N, 1, Float32}))
       end;
  44.674 ns (0 allocations: 0 bytes)
  40.647 ns (0 allocations: 0 bytes)
  16.021 ns (0 allocations: 0 bytes)

julia> let N=128, M=8
           @btime mul!($(zeros(Float32, M)), $(rand(Float32, M, N)), $(rand(Float32, N)))
           @btime $(rand(SMatrix{M, N, Float32})) * $(rand(SVector{N, Float32}))
           @btime $(rand(SMatrix{M, N, Float32})) * $(rand(SMatrix{N, 1, Float32}))
       end;
  65.031 ns (0 allocations: 0 bytes)
  351.633 ns (0 allocations: 0 bytes)
  46.571 ns (0 allocations: 0 bytes)

The equivalent matrix matrix product scales fine, so my guess is the necessary functionality is there, but it needs to be wired up?

Reproduce the reported @btime comparisons for mul!, SMatrix * SVector, and SMatrix * SMatrix at N=32, 64, and 128. Start by tracing the matrix-vector multiplication and generated-code path, then compare its dispatch with the matrix-matrix product. Done means the matrix-vector product no longer shows the reported scaling problem while preserving zero allocations.