GPU Test stableness tracker

List of unstable GPU tests

• pytorch_CycleGAN_and_pix2pix train: https://github.com/pytorch/benchmark/actions/runs/4147131894

• DALLE2_pytorch eval (33% variation): https://github.com/pytorch/benchmark/actions/runs/4147979273
  Reason: eval is dominated by memory copy
  

• LearningToPaint train (115% variation): https://github.com/pytorch/benchmark/actions/runs/4287347886
  Try run 1: https://github.com/pytorch/benchmark/actions/runs/4295256456 (remove CPU tensors)
  Try run 2: https://github.com/pytorch/benchmark/actions/runs/4347585042

• dense121 eval (11.2% variation):

• detectron2_fasterrcnn_r_50_c4 eval (16% variation):

• dlrm eval (15% variation):

• mnasnet1_0 train (17% variation):

• phlippe_densenet train (10% variation):

• pytorch_CycleGAN_and_pix2pix eval (15% variation):

• resnext50_32x4d train (10.7% variation):

• shufflenet_v2_x1_0 eval (14.8% variation):

• timm_vision_transformer train (16% variation): https://github.com/pytorch/benchmark/actions/runs/4147202172
  Fix: https://github.com/pytorch/benchmark/pull/1419

• Yolov3 train: https://github.com/pytorch/benchmark/actions/runs/4145521493

• Yolov3 eval: https://github.com/pytorch/benchmark/actions/runs/4146281655

Start with the unchecked LearningToPaint train item and review its linked GitHub Actions runs, including the two listed attempts and the note about CPU tensors. Compare the remaining unchecked GPU benchmark tests and their reported variation. Done means the unstable tests have been investigated and the tracker items are resolved or updated with evidence.