Fit linear plane in pyxem.signals.BeamShift.get_linear_plane by minimising magnitude variance

Checklist

• implementation steps
• update the Changelog
• mark as ready for review

What does this PR do? Please describe and/or link to an open issue.
In my work, there can be electromagnetic field in the sample areas where I want to fit linear planes for descan removal. However, using the existing algorithm in pyxem.signals.BeamShift.get_linear_plane by minimising the magnitudes through least squares gives poor results.

What is better for my case is assuming the sample has uniform electromagnetic field strengths and therefore uniform deflection magnitudes, then fitting the linear planes from this. This is done by minimising the deflection magnitude variance so that ther is a constant deflection magnitude after fitting. This can be used by setting constrain_magnitude to True:

import hyperspy.api as hs
import pyxem as pxm
import numpy as np

p = [0.5] * 6  # Plane parameters
x, y = np.meshgrid(np.arange(256), np.arange(256))
base_plane_x = p[0] * x   p[1] * y   p[2]
base_plane_y = p[3] * x   p[4] * y   p[5]

# Base descan plane
base_plane = np.stack((base_plane_x, base_plane_y)).T
data = base_plane.copy()

# Add electromagnetic shifts, four domains.
shifts = np.zeros_like(data)
shifts[:128, 128:] = (10, 10)
shifts[:128, :128] = (-10, -10)
shifts[128:, 128:] = (-10, 10)
shifts[128:, :128] = (10, 10)
data  = shifts

s = pxm.signals.BeamShift(data)

# Regular plane fitting gives poor results
s_lp = s.get_linear_plane()
assert not np.allclose(s_lp.data, base_plane.data, rtol=1e-7)

# By minimising magnitude variance you get the original plane back
s_lp = s.get_linear_plane(constrain_magnitude=True)
assert np.allclose(s_lp.data, base_plane.data, rtol=1e-7)

However there are a couple issues. If there are few domains, then there will be many valid planes for this fitting, some very nonsensical. Therefore, I have exposed the initial_values parameter, which you can vary to get different results. Part of the reason is that since we fit the planes together, we have six parameters and therefore the algorithm seems to get stuck in many local minima. See the tests for an example where this is used to fix things. The algorithm is also very affected by noise, so a mask is often necessary where this is severe.

I have also added a check for the mask having a datatype of booleans. It can accept other datatypes silently, where it would most likely give wildly wrong results. So I found it useful to have an explicit check.

Regardless this is a very useful feature for me that has helped with my data-analysis. So I hope we can share it with others!