Support building GPU-capable sherpa-onnx on Linux aarch64. (#1500)

Thanks to @Peakyxh for providing pre-built onnxruntime libraries 
with CUDA support for Linux aarch64.

Tested on Jetson nano b01

sherpa-onnx/csrc/online-cnn-bilstm-model.cc

@@ -47,7 +47,7 @@ class OnlineCNNBiLSTMModel::Impl {
     return {std::move(ans[0]), std::move(ans[1])};
   }
 
-  OrtAllocator *Allocator() const { return allocator_; }
+  OrtAllocator *Allocator() { return allocator_; }
 
   const OnlineCNNBiLSTMModelMetaData &GetModelMetadata() const {
     return meta_data_;