Better stereo combiner for disp predictions

[maxnoe]

Please describe the use case that requires this feature.

Currently we just average the single telescope predictions on the sky.

This is not optimal, since it will happily average a telescope with wrongly reconstructed sign of disp with another with correctly reconstructed sign resulting in a very bad prediction.

There are multiple ways to combine disp predictions for multiple telescopes, e.g. taking the average of the closest positions taking both possibilities of sign into account.

Some where looked at in @LukasNickel Master's thesis.

Describe the solution you'd like

Implement improved combination of telescope-wise disp prediction into subarray prediction.

Additional context

This method seems to provide better performance for low multiplicities than HillasReconstructor.