Fix bug on safe_log and solves discontinuous magnetic fields

[santisoler]

Fix how _safe_log checks if r == |x|. Add the other two shifted coordinates (y and z) as arguments for _safe_log so we can use them to assess that the observation point is inline with a pair of vertices or not. Make the check by evaluating if y == 0 and z == 0 instead of -x == r (with x < 0), since -x and r could be the same value (up to machine precision) while y and z are non-zero (specially when $|x| \gg |y|$ and/or $|x| \gg |z|$). Update calls to _safe_log, update its docstrings, and update docstrings of kernels that mentioned the safe log function. Add tests that catches the bug, and that pass with the bugfix.

Relevant issues/PRs:

Fixes #98

[santisoler]

Running the snippet I pasted in #98 on this branch, I get the following image, showing that the fields are well evaluated with this bugfix:

[santisoler]

@leouieda I'm merging this PR since the tests I wrote verify that the bugfix solves the issue without introducing new errors, and it doesn't impact the performance of the kernels. Feel free to comment if you have any thought!

[santisoler]

Ok, one more check that this bugfixes improves the code.

Here is a snippet that evaluates the _safe_log on a set of observation points:

import numpy as np
import matplotlib.pyplot as plt

from choclo.prism._kernels import _safe_log


def eval_log(easting, northing, upward):
    radius = np.sqrt(easting**2 + northing**2 + upward**2)
    try:
        result = _safe_log(upward, easting, northing, radius)
    except TypeError:
        result = _safe_log(upward, radius)
    return result


spacing = 1e-9
vmax = 1e-5
size = int(2 * vmax / spacing) + 1
easting = np.linspace(-vmax, vmax, size)
northing, upward = 0.0, -50.0

results = np.array([eval_log(e, northing, upward) for e in easting])
results

plt.plot(easting, results)
plt.show()

With the code in main I get the following result:

While with this branch I get:

With this bugfix we solve the discontinuity issue, but we also improve the accuracy of the evaluations of the safe_log.