cudnn_conv_algo_search ignored when set via gpu_execution_accelerator CUDA EP parameters

**Description** When configuring `cudnn_conv_algo_search` for the CUDA execution provider via the `optimization.execution_accelerators` syntax, the setting is silently ignored and `EXHAUSTIVE` is used regardless of the value specified. **Triton Information** Version: 24.10 Container: `nvcr.io/nvidia/tritonserver:24.10-py3` (ORT 1.19.2) **To Reproduce** Model config using the documented EA syntax: ``` platform: "onnxruntime_onnx" input [{ name: "input" data_type: TYPE_FP16 dims: [-1,3,32,256] }] output [{ name: "output" data_type: TYPE_FP16 dims: [-1,73,96] }] optimization { execution_accelerators { gpu_execution_accelerator: [{ name: "cuda" parameters { key: "cudnn_conv_algo_search" value: "1" } }] }} ``` Launch Triton with `--log-verbose=2`. The verbose log shows: ``` CUDA Execution Accelerator is set for 'model' on device 0 with options: ...cudnn_conv_algo_search=EXHAUSTIVE... ``` This occurs with both `value: "1"` and `value: "HEURISTIC"`. The same model using the deprecated syntax applies the setting correctly: ``` parameters { key: "cudnn_conv_algo_search" value: { string_value: "1" } } ``` Verbose log with deprecated syntax: ``` CUDA Execution Accelerator is set for 'model' on device 0 with options: ...cudnn_conv_algo_search=HEURISTIC... ``` **Expected behavior** Setting `cudnn_conv_algo_search` via `optimization.execution_accelerators.gpu_execution_accelerator` should apply as configured. The verbose log should reflect the specified value, not `EXHAUSTIVE`.