报告一些关于 RayIoU 的小错误，希望指标计算细节社区能统一一下（已⭐️）

1. EgoPoseDataset 生成的激光雷达原点可能在场景外，作者意图应该还是想限制在场景内的吧，但是代码里只考虑俯视平面，没考虑到自车的姿态，比如在下坡场景中远处的位置实测是有部分原点在地下或上天的，在我的实验中这里会带来 (0.426182-0.424588) 的 1e-3 级的差别

# if np.abs(origin_tf[0]) < 39 and np.abs(origin_tf[1]) < 39:
if np.abs(origin_tf[0]) < 39 and np.abs(origin_tf[1]) < 39 and (origin_tf[2] > -1 and origin_tf[2] < 5.4):
    output_origin_list.append(origin_tf)

2. calc_rayiou 中要不要也像计算普通 iou 一样显式置 nan，这只对计算迷你验证集上计算指标有影响，完整验证集上不会出现某个真值类从不出现的情况

def per_class_iou(hist):
    result = np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist))
# ----------------------------------------------------------------------------------------------------
    result[hist.sum(1) == 0] = float('nan')
# ----------------------------------------------------------------------------------------------------
    return result

def calc_rayiou(pcd_pred_list, pcd_gt_list, occ_class_names):
....
    iou_list = []
    for j, threshold in enumerate(thresholds):
# ----------------------------------------------------------------------------------------------------
        # iou_list.append((tp_cnt[j] / (gt_cnt + pred_cnt - tp_cnt[j]))[:-1])
        iou = (tp_cnt[j] / (gt_cnt + pred_cnt - tp_cnt[j])
        iou[gt_cnt == 0] = float('nan')
        iou_list.append(iou[:-1])
# ----------------------------------------------------------------------------------------------------
    return iou_list

3. 虽然听着很扯，但是下面这段冗余的代码似乎会导致奇怪的 Bug（当前计算指标的代码里用 torch.cuda.empty_cache() 可以解决此 Bug），删除下面的冗余代码后可以不用 empty_cache

Start by locating the EgoPoseDataset origin-generation code and the calc_rayiou and per_class_iou entry points mentioned in the issue. Reproduce the mini-validation behavior, then investigate the origin bounds, NaN handling for absent ground-truth classes, and the redundant code shown in the report. Done means the reported metric differences and cache-related bug are resolved without requiring torch.cuda.empty_cache().