Summary: Remove tiny-tensorrt dependency and rewrite a new tensorrt inference api. In the new version of trt infer, it can pad the input to fixed batch automatically, so you don't need to worry about dynamic batch size.
