Ray actually should have infinite length, so we can detect camera in a volume
which is bigger that the far clipping of the camera.
This might also give some speedup (wouldn't expect much tho) because we don't
need to re-calculate ray direction and length after every bounce now.