add robust and informative condition for large magnitude projections (tests only off-diagonal)

[renanxcortes]

This PR is closely related to #96. This adds a more robust condition that raises for the user what is actually happening during the calculations.

@ljwolf shared with me an example using cenpy for a GeoDataFrame with a large unit for the centroid and realized that the RelativeClustering measure was behaving unexpectedly. It worked for Phoenix, but not for Tucson (was returning an NaN). However, in that case, even for Phoenix the index might be problematic since numerically the internal calculations are unstable. For Phoenix we got virtually zero for the exp(-d) of the distances:

And for Tucson we were having equal zero (therefore resulting in a NaN index):

The commit of this PR fix this and raises an informative ValueError better and perform a more assertive test than the old one.