Split "geometry" into "direction" and "impact"

[maxnoe]

Please describe the use case that requires this feature.

Some algorithms might not be able to predict both impact and direction (e.g. disp).
Some algorithms might only be able to predict impact (e.g. for monoscopic observations, two independent machine learning models).

Should we split the data model for "geometry" into a "direction" and an "impact" component?