This PR relates to issue #3405. Credit to hoanle for finding the issue.

Background:
OccGridEstimator uses different levels of AABBs to determine where to sample the points from with varying levels of detail.

        # multiple levels of aabbs
        aabbs = torch.stack(
            [_enlarge_aabb(roi_aabb, 2**i) for i in range(levels)], dim=0
        )

But those samples outside the original AABB are not used anyway if we disable the scene contraction:

        if self.spatial_distortion is not None:
            positions = ray_samples.frustums.get_positions()
            positions = self.spatial_distortion(positions)
            positions = (positions + 2.0) / 4.0
        else:
            positions = SceneBox.get_normalized_positions(
                ray_samples.frustums.get_positions(), self.aabb
            )

Since the positions (if distortion is disabled) are taken from a normalized standpoint until the self.aabb parameter is passed in.

Problem:
The problem is that if we begin with the scene_aabb, the outermost AABB, we will not use any of the other resolutions in our enlargement calculations. Therefore, we need to start from the innermost AABB depending on the number of levels.