Support for TensorRTLinear model.

* This PR adds support for Tensor RT.
* There are probably more minor improvements I can make to
`tensorrt.py`, but I think this is a good starting point.
* `freeze_model.py` exports metadata about the model which helps us
reconstruct the TensorRT model.
* Make sure `pycuda.autoinit` is never loaded during training, as it
interferes with TF-GPU (when training).

Test: Tested this on my Jetson Nano.