Fix rrpn proposal validation

Summary:
https://github.com/facebookresearch/detectron2/blob/08f617cc2c9276a48e7e7dc96ae946a5df23af3f/detectron2/modeling/proposal_generator/proposal_utils.py#L104
```python
        if not valid_mask.all():
            if training:
                raise FloatingPointError(
                    "Predicted boxes or scores contain Inf/NaN. Training has diverged."
                )
            boxes = boxes[valid_mask]
            scores_per_img = scores_per_img[valid_mask]
            lvl = lvl[valid_mask]
```

Unlike function **find_top_rpn_proposals**, **find_top_rrpn_proposals** does not include valid masking for level indices.
When the program goes into if condition, it gives an undesirable error says that "shape does not match".

I add a handling code same with **find_top_rpn_proposals**.

Pull Request resolved: #3770

Reviewed By: sstsai-adl

Differential Revision: D33119548

Pulled By: zhanghang1989

fbshipit-source-id: fee677ad6b610deb713d2769901ab8d3d47539c3

Author: nijkah

detectron2/modeling/proposal_generator/rrpn.py

@@ -92,17 +92,23 @@ def find_top_rrpn_proposals(
     for n, image_size in enumerate(image_sizes):
         boxes = RotatedBoxes(topk_proposals[n])
         scores_per_img = topk_scores[n]
+        lvl = level_ids
+
         valid_mask = torch.isfinite(boxes.tensor).all(dim=1) & torch.isfinite(scores_per_img)
         if not valid_mask.all():
+            if training:
+                raise FloatingPointError(
+                    "Predicted boxes or scores contain Inf/NaN. Training has diverged."
+                )
             boxes = boxes[valid_mask]
             scores_per_img = scores_per_img[valid_mask]
+            lvl = lvl[valid_mask]
         boxes.clip(image_size)
 
         # filter empty boxes
         keep = boxes.nonempty(threshold=min_box_size)
-        lvl = level_ids
         if _is_tracing() or keep.sum().item() != len(boxes):
-            boxes, scores_per_img, lvl = (boxes[keep], scores_per_img[keep], level_ids[keep])
+            boxes, scores_per_img, lvl = (boxes[keep], scores_per_img[keep], lvl[keep])
 
         keep = batched_nms_rotated(boxes.tensor, scores_per_img, lvl, nms_thresh)
         # In Detectron1, there was different behavior during training vs. testing.