Following Cooper (2000)

It will make easy to find future multiple EQL gridders by tab
autocompletion.

Replace the distance_to_nearest function with verde.meadian_distance and
add k_nearest as optional argument for the gridder.

This new test function is intended to be run with Numba jit disabled.
A great number of masses will heavily increase the computation time if
Numba jit is disabled.

Because the vertical coordinate points downwards, the vertical component
for the point sources must be increased if we want them to be bellow the
observation points.

Replace the preexisting function to compute distances in Cartesian
coordinates for the distance_cartesian function in forward/utils.py.

Remove BlockReduce: it's not wise to use it on harmonic data where
observation points are at different heights.
Reduce data by discarding random points.

Make clearer that the default position of point masses are bellow each
observation point at a relative depth proportional to the distance of
the nearest neighbours.

Reduce region instead of downsampling the dataset. Add comment
explaining what score is. Increase k_nearest because airborne magnetic
survey. Use EQLHarmonic constructor to set parameters instead of
set_params. Make the two axes plot the same region.

Co-Authored-By: Leonardo Uieda <[email protected]>

Removed the code for variable depth based on neighbor distances.
Use a different anomaly for the example and plot balancing the colormap.
Use deeper sources for smoother grid and appropriate grid spacing.
Simple train-test split instead of CV for quicker and easier to
understand example.

Improve docstrings to make clearer that the relative_depth argument
represent the relative depth at which the point sources will be located
referenced to the elevation of the data points.