Implementation of `densify_and_clone()` differs from what is mentioned in the paper.

[pgarun]

In Section 5.2 - Adaptive Control of Gaussians, it is mentioned, "For small Gaussians that are in under-reconstructed regions, we need to cover the new geometry that must be created. For this, it is preferable to clone the Gaussians, by simply creating a copy of the same size, and moving it in the direction of the positional gradient."

The implementation however, does not move the cloned Gaussians in the direction of their positional gradient. Is there a specific reason for this change in the implementation?