Add rules for sqrt(::Diagonal)

[thomasgudjonwright]

Adding an rrule for sqrt(::Diagonal), relating to a comment in an issue discussing structural vs natural Tangents

[mzgubic]

The issue is that the rand_tangent treats Integers as non differentiable, i.e.

julia> rand_tangent(Diagonal([1,2, 3.0]))
3×3 Diagonal{Float64, Vector{Float64}}:
 -2.21   ⋅     ⋅ 
   ⋅    2.4    ⋅ 
   ⋅     ⋅   -8.98

julia> rand_tangent(Diagonal([1,2,3]))
NoTangent()

[mzgubic]

@mcabbott are you happy to approve the current state? I pushed commits so it's good if you approve as well. We couldn't quite figure out what went wrong with Tom's local setup but everything works when he does a fresh git clone

[mzgubic]

LOL, okay, so after pushing the rand(3) to the tests they failed on nightly as well. I think it's that FiniteDifferences is doing some sort of step finding and accidentally calls sqrt of a negative number. Multiplying the rand() by 10 worked locally, but it's probably better to just fix them to some fixed matrix to avoid stochastic failures.

The weird part is that I could only reproduce locally by running the test suite (and not copy pasting the tests) so it looks like this was some kind of interaction with the JuliaInterpretter