Change convention for complex differentiation

[sethaxen]

This issue tracks the necessary changes after https://github.com/JuliaDiff/ChainRulesCore.jl/issues/167 is merged:

• ChainRulesCore:
  • [x] Update ChainRulesCore.@scalar_rule to conjugate the partials in rrule. (breaking change) (https://github.com/JuliaDiff/ChainRulesCore.jl/pull/170)
• ChainRulesTestUtils:
  • [x] Bump ChainRulesCore compat (https://github.com/JuliaDiff/ChainRulesTestUtils.jl/pull/44)
  • [x] Update ChainRulesTestUtils.test_scalar to conjugate the derivative when testing rrule. (breaking change)(https://github.com/JuliaDiff/ChainRulesTestUtils.jl/pull/44)
• ChainRules:
  • [x] Add missing tests for rules that take complex arguments (#214, #215)
  • [x] Bump ChainRulesCore, ChainRulesTestUtils, and FiniteDifferences compats (breaking change) (#198)
  • [x] Update non-scalar rrules to follow the new convention (breaking change).
• Zygote:
  • [x] Bump ChainRules compat (https://github.com/FluxML/Zygote.jl/pull/721)
  • [x] Update Zygote's adjoints to stop conjugating to/from ChainRules (https://github.com/FluxML/Zygote.jl/pull/721)

I have local implementations of most of these already.

[willtebbutt]

Happy to help out anywhere that it would be useful, but you're probably the best placed to deal with these.

[oxinabox]

Update Zygote's adjoints to stop conjugating to/from ChainRules (breaking change)

This is not breaking for Zygote. Its a nonbreaking feature to change what version of ChainRule's it supports

[sethaxen]

Good point, thanks. Yes, if Zygote bumps the compat and stops conjugating to/from ChainRules at the same time, this is not a breaking change. I have updated.