Replace mul! by broadcasting in exp!

[jishnub]

This reduces TTFX by a bit, by avoiding the *ₛ broadcast and inlining the result instead. This also makes the operation easier to read.
TTFX improvement:

julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time LinearAlgebra.exp!(A);
  3.536534 seconds (5.32 M allocations: 278.674 MiB, 8.06% gc time, 100.00% compilation time) # master
  3.132689 seconds (4.88 M allocations: 255.420 MiB, 1.93% gc time, 99.99% compilation time) # This PR