efficiency improvements in autograd derivatives

[tylerflex]

Orders of magnitude improvement to autograd gradient handling, especially for large field data and many structures.

Achieved through following:

1. assuming sorted in field data interpolation for adjoint gradients.
2. not deep copying field data containing DerivativeInfo objects when passing around.

Still work in progress as I need to figure out where else I could apply these fixes.

[yaugenst-flex]

This is great! I guess the only thing I wonder is that derivative_info contents are never mutated? Basically after this we need to make sure that never happens, but seems like that shouldn't be the case anyways?

[tylerflex]

it should never be the case. the derivative info is basically just a container to avoid having to pass Ex_fwd, Ey_fwd, etc into all of the derivative computation methods and then interpolate into them. We'd never mutate them