The value of `hypervolume_plus` when removing the last point

[nikohansen]

is computed to -inf:

        if hasattr(self, '_hypervolume'):
            self._hypervolume_plus = self._hypervolume if self._hypervolume > 0 else -inf

However, this does not seem to agree with the definition of the hypervolume_plus property:

        `hypervolume_plus` equals to the hypervolume when the archive is
        nonempty, otherwise it is the smallest Euclidean distance to the
        hypervolume area (AKA reference domain) times -1 of any element
        that was previously added but rejected because it did not dominate
        the reference point.

It is probably best to recover the last negative value before the HV became positive? If we change ._hypervolume_plus only as long as the archive remains empty, see #33, this would be trivial to implement. The caveat is that some current code could rely on using _hypervolume_plus instead of hypervolume_plus at some places?