Fix the KDE bandwidth in the hierarchical analysis

When the samples get converted to KDEs in the hierarchical analysis, the KDE uses a double version of the samples also containing a copy of the samples reflected about zero. However, this means that when the KDE bandwidth get calculated (in particular for strong signals peaked well away from zero) the bandwith can be much too large, resulting in KDEs that are far broader than the underlying distribution. This MR fixes this by using a bandwidth equivalent to that calculated from the original unreflected samples.