[PyTorch] GPUOpt Runtime: validated CUDA Graph reuse for faster LLM decoding

Part of

Cross-Architecture Performance Modeling Trends