Simple SV fit improvements
This MR makes a few simple changes to the SV fitter to speed up the fit and minimize global memory accesses. This includes the change from !616 (closed). In addition, the SV POCA calculation is moved to the track filter. As a result, the final SV container will contain only valid vertices.
This MR also fixes math errors in the calculation of the SV covariance matrix and FD chi2. These errors were leading to SV chi2 < 0 and incorrect FD chi2s.
The default IP chi2 preselection for secondary vertices is loosened from 9 to 4 to align with the preselection used to train the TwoTrackMVA classifier. In addition, the FD chi2 saved for each SV is now the minimum FD chi2 w.r.t. all PVs, in line with what was used to train the TwoTrackMVA classifier.
The net result is a ~25% decrease in selected Bs -> phi phi events (about 400/1000 -> 300/1000) by the TwoTrackMVA and a ~25% decrease in minimum bias rate (about 750 kHz -> 550 kHz) at the same classifier threshold.
Solves #252 (closed)