Add a new program for benchmarking based on a TensorRT GoogleNet sample

[psyhtest]

Create a new program for benchmarking inference based on /usr/src/gie_samples/samples/sampleGoogleNet (GIE path on L4T), so we can do our usual batch size and model exploration.

In the source code they say:

       // run inference with null data to time network performance
        timeInference(engine, BATCH_SIZE);

(meaning that no weights file is needed?)

So this should be like caffe time (whereas issue #1 is like caffe test).

[psyhtest]

The new tensorrt-time program is now complete. It supports different batch sizes, floating-point precision (fp16, fp32) and caching of created models (based on the batch size, precision and TensorRT library version). I will work on scripting the benchmarking process next.