[RANS] Make calculation of velocity gradients more general and robust

When considering a domain that features a pointy tip like illustrated below, it might be the case that the cell at the tip does not have any vertical or horizontal neighbors. Right now, the calculation of the velocity gradient for RANS seems to produce nan values for this situation.

+++++++++++++   
+++++++++++  
++++++++  
++++++  
+++  
+   

One could furthermore think of a more general gradient calculation tool that might by helpful in other cases.

Robustness issue addressed by !1130 (merged),
the generalization could be done on a dumux day.