Update the signs in the documentation and in the implementation of each
kernel. Also make the kernels to return zero on singular points (not on
vertices) so when computing on lines shared with the edges are not falling
in division by zero errors and the limits for those cases are well
resolved.

Singular points for the grad components should be the same as the ones
for the magnetic field components: interior points, edges and vertices.

Reduce repeated code by dumping the evaluation of the kernels on its own
private function.

Fix minus sign applied to a list and set an atol so we can compare
values close to zero.

Test if evaluation of the mag grad components on singular points return
nans.

This way I quickly verified that the mathematical expressions in the
docstrings match the implementations.