Inference without Horovod (TRI-ML#33)

DGP Multi-Cam + Velocity Loss + FP16 Inference + PackNetSlim (TRI-ML#30)

* Support for multi-camera loading on the DGP dataset
* Support for fp16 at inference time
* Velocity loss (see VelSupModel)
* PackNetSlim01 (faster version of PackNet)

First release of training code (TRI-ML#11)