pass inference accuracy check for detectron2_fcos_r_50_fpn (#108328)
We need a higher tolerance to pass the inference accuracy check for detectron2_fcos_r_50_fpn .
Command:
```
python benchmarks/dynamo/torchbench.py --backend inductor --bfloat16 --accuracy --only detectron2_fcos_r_50_fpn --disable-cudagraphs --inference
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108328
Approved by: https://github.com/jansel
