Apply two methods of latency measurements in the inference

[alexanderwendt]

The first measurement method is like the benchmark_app to jst run through networks without any images. The second method is to load real images and run through the networks with them, i.e. load, infer, load, infer.

It shall be applicable to as the measurement method should be set as an option, either to measure both ways or only the benchmark_app way: tf2oda_inference_from_saved_model.py