Reapply "import Tensorrt not in compile time in order to dynamically load tensorrt only if nvidia gpu is present"

This reverts commit cf01e5d952.

Azaion.Inference/tensorrt_engine.pyx

@@ -56,7 +56,7 @@ cdef class TensorRTEngine(InferenceEngine):
             raise RuntimeError(f"Failed to initialize TensorRT engine: {str(e)}")
 
     @staticmethod
-    cdef get_gpu_memory_bytes(int device_id):
+    def get_gpu_memory_bytes(int device_id):
         total_memory = None
         try:
             pynvml.nvmlInit()
@@ -73,7 +73,7 @@ cdef class TensorRTEngine(InferenceEngine):
         return 2 * 1024 * 1024 * 1024 if total_memory is None else total_memory # default 2 Gb
 
     @staticmethod
-    cdef get_engine_filename(int device_id):
+    def get_engine_filename(int device_id):
         try:
             device = cuda.Device(device_id)
             sm_count = device.multiprocessor_count
@@ -83,7 +83,7 @@ cdef class TensorRTEngine(InferenceEngine):
             return None
 
     @staticmethod
-    cdef convert_from_onnx(bytes onnx_model):
+    def convert_from_onnx(bytes onnx_model):
         workspace_bytes = int(TensorRTEngine.get_gpu_memory_bytes(0) * 0.9)
 
         explicit_batch_flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
@@ -112,13 +112,13 @@ cdef class TensorRTEngine(InferenceEngine):
             constants.log('conversion done!')
             return bytes(plan)
 
-    cdef tuple get_input_shape(self):
+    cpdef tuple get_input_shape(self):
         return self.input_shape[2], self.input_shape[3]
 
-    cdef int get_batch_size(self):
+    cpdef int get_batch_size(self):
         return self.batch_size
 
-    cdef run(self, input_data):
+    cpdef run(self, input_data):
         try:
             cuda.memcpy_htod_async(self.d_input, input_data, self.stream)
             self.context.set_tensor_address(self.input_name, int(self.d_input))  # input buffer