fixpoint AD interface design

[GiggleLiu]

The separation into two argument-tuples was meant to mirror the two different partial derivatives that need to be taken in an iterative adjoint (as in your paper). Same for the fixed-point procedure - if we want to add the fixedpoint backpropagation we'll have to change these things back at some point. Not a big fan of that change but given the time it might be ok.

Originally posted by @under-Peter in #5