Support ONNX models in TensorRT backend

[hzyhhzy]

This PR enables the TensorRT backend to support ONNX models, facilitating experimentation with new architectures such as Transformers. This implementation references the work of @yehu3d.

• Dual Model Support: Supports both .onnx models and traditional .bin.gz models. The format is automatically detected based on the file extension.

• Model Export: The ONNX export script is available here: export_onnx.py.

  • Note: This script is designed for the KataGo_Transformer repository. Minor modifications may be required to adapt it for the official KataGo's PyTorch code.

• While different board sizes are supported, there is a known limitation where pos_len is fixed during the ONNX export. Consequently, TensorRT inference is restricted to using nnXLen and nnYLen values that match the fixed pos_len. So compatibility with different board sizes is currently handled via masking, which incurs a certain degree of performance loss.