MLIR-fusion seems to corrupt CUDA graph replay in Gemma4 e2b

[bmarimuthu-nv]

MLIR elementwise kernels currently corrupt piecewise CUDA graph replay for Gemma4 E2B.

Steps to reproduce:
examples/auto_deploy/model_registry/configs/gemma4_e2b.yaml
enable mlir_elementwise_fusion
and run the model e2e.

  mlir_elementwise_fusion:
    enabled: true

[nvchenghaoz]

Will take a look once the PR is merged.