Disparate performance in diff methods

First of all, thank you for this great package! I'm a relatively new Julia user and I was using Static Arrays for some polynomial math that needs to use diff. I noticed that the benchmark speeds are very different for the two diff methods. MWE:

julia> A = @SMatrix randn(3,20);

julia> @btime diff($A,dims=$2);
  3.299 μs (2 allocations: 1.06 KiB)

julia> @btime diff($A,dims=$(Val(2)));
  37.283 ns (0 allocations: 0 bytes)

Start by reproducing the two diff benchmarks from the issue with a 3×20 @SMatrix and compare the dims=2 and dims=Val(2) paths. Trace the diff implementation used by StaticArrays and identify why their allocations and timings differ; done means the discrepancy is explained and the intended behavior is covered by a regression test or benchmark.