Refactor the forward backward classes

[tkoskela]

SOPT has several "solvers" that take the operators f_gradient and g_proximal as arguments and several definitions for those operators using different norms. We could implement the learned model as a "norm" and pass it in using the current interface with minimal modifications to the rest of the code. Note that the code on lines 234-236 of ForwardBackward.h is specific to the L1 norm and needs to be put in a function.

Currently the definitions of different norms have a lot of copy-pasted code that should be merged while we add the learned model. The current norms for the forward backward solver are imaging_forward_backward.h and l2_forward_backward.h. PR #256 adds another one to the primal dual solver.

Create a new class that combines the classes in imaging_forward_backward.h, l2_forward_backward.h (including some of the clean-up in #256)

Once this is done, we can proceed with

• [x] #298
• [x] #299

[tkoskela]

Notes from pair programming session:

• The tests in forward_backward.cc are only testing that the properties of the class have the correct type. We may need some more checks before we start changing things
• How can we run the forward-backward algorithm on an image?
• Do we want to maintain l2_forward_backward, as it's not being used, or just cut it?
• Macros in imaging_forward_backward.h. Do we want to keep them or rewrite in plain c++ for clarity?