Flexible metadata

[reckart]

The protocol right now seems to have a fixed set of metadata which is forwarded to the classifier as arguments e.g. to the predict call. If ariadne is used outside the context of INCEpTION, information such as the project ID or document ID or user may not be required. So I believe it should be considered to make the metadata more flexible:

• allow basically an arbitrary JSON object to be used as metadata
• pass that object as a single argument to the relevant methods (train / predict)
• define a few well-known keys that can / should be used in the metadata object but that should not always be mandatory