Replace broadcast by loop in LinearAlgebra.exp! for StridedMatrixes

[jishnub]

Since the broadcasts are carried out over StridedMatrixes of the same size, we may replace these with explicit loops over the arrays. The loop is much simpler to compile, and this reduces TTFX by quite a bit:

julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time LinearAlgebra.exp!(A);
  3.253866 seconds (4.85 M allocations: 254.398 MiB, 6.92% gc time, 100.00% compilation time) # nightly v"1.12.0-DEV.558"
  1.819190 seconds (3.76 M allocations: 194.803 MiB, 3.01% gc time, 99.99% compilation time) # this PR