Vp-suite

- Overhauled API and many fixes for usability
- Improved visualization and logging capabilities
- Improved measures (metrics and losses)
- Added new Dataset: `MMNIST on the fly` (`MMF`)
- Set up github page with complete documentation
- Added option to specify save location for models, datasets and logs

- New models that match their official implementations (ConvLSTM/TrajGRU - powered Encoder-Forecaster model from [this repo](https://github.com/Hzzone/Precipitation-Nowcasting; PredRNN++ from [this repo](https://github.com/thuml/predrnn-pytorch))
- New model blocks
- Automatic implementation match tests and better documentation of which models, model blocks, datasets and measures are available

New Datasets:
- Physics101
- Human 3.6M
- KITTI-Raw
- Caltech Pedestrian

- Refactored `Trainer` and `Tester` into `VPSuite`: Multiple models and datasets can now be loaded and flexibly used for testing and training.
- Added basic tests for all models, datasets and measures (metrics and losses).
- Updated README to reflect changes.

Now with correct pathing: install directly from test-PyPi and use from any working directory!

A folder named `vp-suite` will be created where you are, containing downloaded data and run results.

The following datasets are now automatically downloaded and prepared if `Trainer`'s `load_dataset()` is called without providing a path to valid prepared data:
- Moving MNIST
- KTH Actions
- BAIR Robot pushing