Interaction of Complex{Dual} with rrule for ^ on Julia ≤ 1.5

[mcabbott]

In Julia 1.5 and earlier, complex dual numbers did not work well. This leads to the following interaction between the log(complex(x))) in the rrule for ^, and Zygote's forward-over-reverse Hessian calculation:

julia> hessian(x -> x[1]^x[2], [3,4])
ERROR: StackOverflowError:
Stacktrace:
 [1] Complex at ./complex.jl:12 [inlined] (repeats 2 times)
 [2] float at ./complex.jl:1016 [inlined]
 [3] log(::Complex{ForwardDiff.Dual{Nothing,Float64,2}}) at ./complex.jl:583 (repeats 79984 times)

It did not happen prior to #513 (v1.11.4) because this rule had a different special case for real x and real p, which I think got some of the edge cases wrong. Perhaps we could do both. But I'm not sure how much work is justified in fixing this. Found in CI & avoided here.

[oxinabox]

maybe we could check if Julia is VERSION < v"1.6" and then do the bad thing that might get some edge cases wrong, just so we can avoid making Zygote's CI for 1.3 and 1.5 fail.

[mcabbott]

Zygote's CI will skip that test as soon as FluxML/Zygote.jl#1044 is merged.

[oxinabox]

For my reference this is fixed in 1.6 by JuliaLang/julia#36030

[mcabbott]

Closing since the package only supports 1.6 now.