🐛 [Bug] group_norm implementation may slows down SD vae

[cehongwang]

Bug Description

We noticed VAE Perf Gap b/w Torch-TRT and ONNX-TRT. ONNX-TRT is about 30-40% faster than Torch-TRT. We found that for group_norm layer, ONNX-TRT is using a workaround provided by TensorRT team, InstanceNormalization. Torch-TensorRT uses add_normalization for direct conversion.

[HolyWu]

Should be fixed by #3719.