Improve the starting position of vertex fit for T-tracks by using the yz straight line intersection

Currently for vertex fitting T-track pairs, the tracks are extrapolated with the Runge-Kutta extrapolator in the first iteration of the vertex fit only.

It does not make sense to extrapolate to the 3D linear point of closest approach (POCA) because the tracks are curved in the x-z plane. This introduces a large bias in z and significantly reduces the mass resolution.

However, the tracks are approximately straight lines in the y-z plane. Instead of extrapolating tracks to the x,y,z point of closest approach, we can extrapolate to the z-position of the y-z straight line intersection of the two tracks, which is usually much closer to the true vertex position.

This MR introduces this change.

This significantly improves the vertex quality, mass resolution, mass bias and z-bias compared to the POCA starting position. There is around a 10-15% drop in efficiency in the fiducial region, mostly concentrated closer to the UT.

The following plots show improvements for signal candidates in Lb -> Jpsi Lambda (Lambda -> p pi) signal Monte Carlo.

this is used by T-track lines here https://gitlab.cern.ch/lhcb/Moore/-/blob/master/Hlt/Hlt2Conf/python/Hlt2Conf/lines/bandq/hlt2_bandq.py