Reduce box_detections_per_img for vision_maskrcnn (#115487)

Summary:
This fixes a failure on the [perf dashboard](https://hud.pytorch.org/benchmark/compilers) with `--amp` mode.  I believe boxes 5 and 6 were getting swapped.  The existing comment explains the issue.

Before
```
$ ./benchmarks/dynamo/torchbench.py --training  --accuracy --no-translation-validatio --amp --backend=inductor --disable-cudagraphs --only vision_maskrcnn
...
[2023-12-09 13:21:27,292] torch._dynamo.utils: [ERROR] RMSE (res-fp64): 0.00171, (ref-fp64): 0.00054 and shape=torch.Size([256, 256, 3, 3])
[2023-12-09 13:21:27,292] torch._dynamo.utils: [ERROR] Accuracy failed for key name backbone.fpn.layer_blocks.2.0.weight.grad
fail_accuracy
```

After
```
$ ./benchmarks/dynamo/torchbench.py --training  --accuracy --no-translation-validatio --amp --backend=inductor --disable-cudagraphs --only vision_maskrcnn
...
pass
```

X-link: pytorch/pytorch#115487
Approved by: https://github.com/yanboliang

Reviewed By: osalpekar

Differential Revision: D52062336

Pulled By: jansel

fbshipit-source-id: bb900b583113c9bda5c13990ea643827b06c2211

userbenchmark/dynamo/dynamobench/torchbench.py

@@ -417,8 +417,8 @@ def load_model(
             # comparison hard with torch.compile. torch.compile can cause minor
             # divergences in the output because of how fusion works for amp in
             # TorchInductor compared to eager.  Therefore, instead of looking at
-            # all the bounding boxes, we compare only top 5.
-            model_kwargs = {"box_detections_per_img": 5}
+            # all the bounding boxes, we compare only top 4.
+            model_kwargs = {"box_detections_per_img": 4}
             benchmark = benchmark_cls(
                 test="train",
                 device=device,