Add ONNX opset 23 RMSNormalization operator support

[AditiThirdEye]

Implements RMSNormalization operator for TensorRT ONNX parser, enabling deployment of modern transformer architectures (LLaMA, Mistral, etc.) that use RMSNorm instead of LayerNorm.

Implementation details:

• Computes Y = (X / sqrt(mean(X^2) + epsilon)) * scale
• Supports FP32, FP16, and BF16 data types
• Handles axis attribute for normalization dimensions
• Supports epsilon and stash_type attributes per ONNX spec

Changes:

• onnxOpImporters.cpp: Add RMSNormalization importer using TensorRT primitive operations (ElementWise, Reduce, Unary)
• onnxOpCheckers.cpp: Add empty checker for RMSNormalization
• docs/operators.md: Add RMSNormalization to supported operators matrix
• onnx_backend_test.py: Include RMSNormalization tests

Fixes onnx/onnx-tensorrt#4639 (via NVIDIA/TensorRT#4639)

[AditiThirdEye]

@kevinch-nv @yuanyao-nv Could you please review this PR when you have a chance?

This adds ONNX opset 23 RMSNormalization support, enabling deployment of modern LLM architectures (LLaMA, Mistral, etc.) that use RMSNorm.

Related issue: NVIDIA/TensorRT#4639

Thanks!

[yuanyao-nv]

Thanks for your contribution. RMSNorm support will be available in the 10.15 release, please stay tuned.

[AditiThirdEye]

Thanks @yuanyao-nv! Just to clarify - will this PR be considered for the 10.15 release, or is there already an internal implementation in progress?
Happy to address any feedback if you'd like to use this PR, or close it if there's already work underway internally.

[yuanyao-nv]

It's the latter. There is already an internal implementation ready to be released in 10.15.