ScaledPlan backwards rule

[gaurav-arya]

The current adjoint for FFT plans in Zygote doesn't handle ScaledPlans. See for example:

JuliaMath/FFTW.jl#182
SciML/OperatorLearning.jl#11

This issue mostly arises because the inverse plan of e.g. an FFT makes use of ScaledPlan. This rule adds a ScaledPlan rule, and adds more comprehensive tests which explicitly test every possible forward and inverse plan (and therefore test for the ScaledPlan issue that was not being handled previously).

Also, it seems like backwards rules for ffts are being moved into AbstractFFTs.jl using ChainRules. However, there are no backwards rules there for FFT plans yet, see JuliaMath/AbstractFFTs.jl#63. (The same issue holding that back -- that the type of an FFT plan does not convey enough information for the backwards rule for real FFTs -- leads to a few errored tests in Zygote, which are manually skipped in gradcheck.jl). But until that happens, it seems like the rules in Zygote are the only ones that handle FFT plans to at least some extent, so this patch should be helpful until plans are properly handled in AbstractFFTs.jl.

[devmotion]

Just a heads up: The rules for rfft etc. are incorrect in Zygote but implemented correctly in AbstractFFTs (#1158).

[gaurav-arya]

I see. I think that this PR, which partially handles FFT plans, still make sense, do you agree? But perhaps this PR / a separate PR should get rid of all the non-plan FFT rules in Zygote, what do you think?

[gaurav-arya]

Closing because we should hopefully have ChainRules for plans somewhat soon, after which this rule can just go.