Workaround for performance penalty of splatting a number

[tkf]

This PR adds a workaround for the performance penalty of splatting a number (#29114) which is problematic in mul! performance.

I don't know the time-frame of the solution of #29114 but it may make sense to have this workaround until we have a builtin solution for #29114?

close JuliaLang/LinearAlgebra.jl#661; cc @dkarrasch @KristofferC

Quick benchmarks

Before (9a8b2fd):

julia> x = rand(10); y = similar(x); λ = rand();

julia> @btime mul!($y, $λ, $x);
  938.111 ns (30 allocations: 640 bytes)

julia> @btime $y .= $λ .* $x;  # as a reference
  7.870 ns (0 allocations: 0 bytes)

After (aeaa464):

julia> @btime mul!($y, $λ, $x);
  9.506 ns (0 allocations: 0 bytes)

The difference to the broadcasting may be the @simd macro.