Separating uses of CoDual into distinct types

[willtebbutt]

Currently CoDual is used for two purposes:

1. to pair a primal and its fdata, and
2. to pair a primal and its tangent.

Having two uses for the same type, only distinguishable by the context, is not good design. It was laziness on my part when I introduced fdata / rdata.

A better design would be to have two distinct types: PrimalFDataPair and PrimalTangentPair. These will be abbreviated to PFPair and PTPair respectively (otheriwse the code is going to become a lot more verbose).

Introducing this distinction will also substantially reduce the chances of bugs in interface code (value_and_gradient!!, DifferentiationInterface.jl's wrappers, etc) -- currently it's very straightforward to accidentally make a CoDual containing a primal + tangent, when you ought to have produce a CoDual containing a primal + fdata.

There are a couple of functions which will need to change -- for example zero_fcodual and zero_codual will become zero_primal_fdata_pair and zero_primal_tangent_pair. More verbose names are acceptable for these, as they do not appear in literally every rrule!! signature. There are probably some more that will appear during refactoring.