Changes from #908 lead to `NaN`s in AD jacobians for methods using `norm2` on `StaticArray` types

[ferrolho]

I recently opened this issue on ForwardDiff.jl: JuliaDiff/ForwardDiff.jl#547. @KristofferC identified #908 as the PR that introduced the breaking changes; see his comment here: #908 (comment). @mateuszbaran said this is because of a missing check in StaticArrays.jl for norm2, which LinearAlgebra already does; see his comment here: #908 (comment).

I am not familiar with the internals of StaticArrays.jl, so any help with this is be much appreciated. Thank you!