Support for NasBench101

[MLRichter]

Is your feature request related to a problem? Please describe. For Neural Architecture Search it could be highly beneficial to have some sort of heuristic to detect unproductive layers. Therefore some interface for importing NASBench-models into this system would be beneficial, so that layers can be removed automatically if desired by the NAS-Algorithm.

Describe the solution you'd like Primarily, an encoder is needed which consumes a network definition in the format provided by NASBench101. The next step would be to add some quality of life functions, which provide the unproductive layers of the model in question in a way that allows for easy automated architectural adjustments.

Additional context Idea still needs to be fleshed out