GeneralizedRCNNWithTTA with horizontal and vertical flip

## 🚀 Feature
I propose to add to GeneralizedRCNNWithTTA not only vertical flip, that I can trigger with cfg.TEST.AUG.FLIP, but even horizontal flip, with another parameter, i.e. cfg.TEST.AUG.HFLIP and cfg.TEST.AUG.VFLIP

## Motivation & Examples

I created by myself a CustomDatasetMapperTTA, that achieve H and V flip:

```
flipH = RandomFlip(prob=1.0, horizontal=True, vertical=False)
aug_candidates.append([resize, flipH]) # resize + flipH

flipV = RandomFlip(prob=1.0, horizontal=False, vertical=True)
aug_candidates.append([resize, flipV]) # resize + flipV

```
but when I run instance segmentation, I get wrong masks, because of the assumption [here in _reduce_pred_masks](https://github.com/facebookresearch/detectron2/blob/f50ec07cf220982e2c4861c5a9a17c4864ab5bfd/detectron2/modeling/test_time_augmentation.py#L298)

Start in detectron2/modeling/test_time_augmentation.py at _reduce_pred_masks and inspect how flip assumptions affect augmented mask predictions. Trace the existing cfg.TEST.AUG.FLIP behavior and the referenced CustomDatasetMapperTTA examples, then define what HFLIP and VFLIP should produce. Done means horizontal and vertical flip options work without incorrect instance masks.

Written by the indexing model from the issue text.