use rrules even when all the arguments are types

[oxinabox]

Fixes this one of SciML/SciMLBase.jl#69 (comment)
(cc @ChrisRackauckas)
Problem was that our resolving order was:

1. ZygoteRules definition (directly overloading _pullback
2. Check if should be ignored based on types of input
3. ChainRules definition
4. generate code via decomposition.

But what it should have been, and what this PR changes it to is:

1. ZygoteRules definition (directly overloading _pullback
2. ChainRules definition
3. Check if should be ignored based on types of input
4. generate code via decomposition.

Also I think the new code is clearer, all chainrules stuff is together, and all the stuff where Zygote is working it out is together.

[oxinabox]

@ChrisRackauckas can you click approve?

[ChrisRackauckas]

How much money do you have on you?

[DhairyaLGandhi]

Otherwise is fine, I think. I'm a bit iffy on whether we should be leaning on CR since we might want to preferentially dispatch to methods in Zygote first for different reasons.

[oxinabox]

preferentially dispatch to methods in Zygote first for different reasons.

This is after that has already happened before we get to this point.

Because @adjoint defines a _pullback that will get hit rather than hitting the generated function.