Add wavefield preconditioner

[jfowkes]

The secret sauce of the LSQ-ML algorithm seems to be what they call average update directions:

In reality this is just scaling the ML update directions by 1 / Lipschitz constant for the exit wave error as shown by Andy Maiden et al in "Further improvements to the ptychographical iterative engine". Note that 1/L is known to be the optimal step size for gradient descent on convex functions in the math optimization literature:
https://math.stackexchange.com/questions/3587312/the-biggest-step-size-with-guaranteed-convergence-for-constant-step-size-gradien

This PR adds an option to ML called "wavefield_precond" that enables this "average update preconditioner". The following results on Moonflower suggests that this is highly effective at accelerating ML as shown in the images below.

Vanilla ML (all parameters default, 200 iterations):

ML with wavefield_precond=True (all other parameters default, 200 iterations):

[pierrethibault]

You need to take the square root of the sum, not the sum of the square roots.

[jfowkes]

You need to take the square root of the sum, not the sum of the square roots.

How stupid of me, thank you! This now works as expected :)

[jfowkes]

Further testing shows that this "wavefield preconditioner" is also highly effective at accelerating ML on real data, as demonstrated below on the small nanogold dataset from the ptypy tutorials:

Vanilla ML (all parameters default, 200 iterations):

ML with wavefield_precond=True (all other parameters default, 200 iterations):

Really impressive to see this level of improvement on ML without any l2-regulariser or smoothing-preconditioner!

[jfowkes]

@daurer now we just need the four CUDA kernels ob_update_ML_wavefield.cu, ob_update2_ML_wavefield.cu, pr_update_ML_wavefield.cu, pr_update2_ML_wavefield.cu to compute the object/probe fluences. I would add these based on the non-wavefield ones but I don't understand how the CUDA indexing works in the kernels.

Then I can add some tests for the cupy and pycuda kernels and we should be good to go.