InteractionBounds are stored as two positions which will fail when one point is at infinity

[austinschneider]

The InteractionBounds are currently represented as a std::tuple<Vector3D, Vector3D> pair of positions.

InteractionBounds are used to compute interaction- and column-depth integrals and thus need to be well defined.

This will process will certainly fail if one of the two positions is a point at infinity, because the direction information is lost.
The solution to this may be to implement a dedicated Ray class.

However, even with a dedicated Ray representation there may still be an issue computing the normalized position probabilities with an unbounded ray. This is something we should double check and potentially implement a special case for.

The main defense against this right now is that we (almost?) always use the Path::ClipToOuterBounds() function during the vertex injection and are initializing our Path objects with an initial position, direction, and length. Which means as long as the vertex injection happens with a valid path through the geometry, it should produce finite InjectionBounds. The trouble will come if we have a path that does not intersect with the geometry, or if some position distribution does not use this same mechanism to calculate its bounds.