Fix FP16 models having poor accuracy

Bugfix for batch demo: add padding if remaning number of frames is not equal to batch size

Make ONNX models compatible with ONNXruntime's TensorrtExecutionProvider