Torch-liberator

Fixed
* Work around networkx removal of OrderedDiGraph

Changed:

* Moved implementation of the backend maximum common subtree algorithm to `https://github.com/Erotemic/networkx_algo_common_subtree`.
* The weight mapping is now returned by the partial state loading and pretrained forward
* Changed defaults to use the isomorphism algorithm by default and other tweaks.

Changed

* Using scikit-build and CMake to build cython modules and publishing wheels
instead of hacky autojit solutions.

Added
* Ported the maximum common subtree isomorphism building associations in `load_partial_state`.

Added
* Added `cli` for manually making deployed zip files

Changed
* Removed Python 2.7 and 3.5 Support.
* The first part in the deployed name is now based on the netharn run "name"
rather than the name of the model class

Fixed
* Support new netharn checkpoints dir

Added
* Added `extract_snapshot` method to `DeployedModel`