[inductor] Update dynamic vision_maskrcnn graph break baselines (#194630)

Human Context: fixing inductor periodic

## Summary

`Fix Dynamo tracing of compile support checks` (#185528) reduced
`vision_maskrcnn` graph breaks from 18 to 15 for inference and from 38 to 33
for training. It updated the regular TorchBench baselines but missed their
dynamic CUDA and ROCm counterparts.

Because `check_graph_breaks.py` intentionally rejects improvements until the
baseline is updated, this has caused repeated Inductor periodic failures. This
PR aligns all four dynamic baselines with the observed counts.

## Test Plan

```text
python benchmarks/dynamo/check_graph_breaks.py --actual /tmp/pytorch-work/agent_space/pt2-oncall/run-32730575308/rocm-reports/test-reports/inference_torchbench.csv --expected benchmarks/dynamo/ci_expected_accuracy/rocm/dynamic_inductor_torchbench_inference.csv
spin quicklint
lintrunner -a
```

The checker reports `vision_maskrcnn PASS` against the failed ROCm periodic
artifact.

This PR was authored with assistance from an AI coding assistant.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/194630
Approved by: https://github.com/karthickai