`DirectNavigator` can fail in rare cases with `CylinderSurfaces`

After diving into physmon changes of https://github.com/acts-project/acts/pull/4694 and https://github.com/acts-project/acts/pull/4887 I believe we have a rare edge case with the `DirectNavigator` and `CylinderSurfaces` if the trajectory deviates from the initial one which estimated the surface sequence.

The problem seem to come from the fact that we disambiguate intersection solutions on the fly due to the usual `CylinderSurfaces` problem and the fact that the surface order can change with infinite bounds. In some cases the correct cylinder intersection will then be behind the current position and be rejected and the second solution is accepted. In such a case the propagation will practically aim for the other side of the detector and the fit will not be successful.