Changed the build-point of the feature factory. It works now together with the LBN as intended

As described in Issue #1, the keras layer does not work as intended. The feature factory caches the tensors and thus any new call leaves the connection to the first input tensor untouched. Since no_cache is only an internal option, I rebuild the build/call function, in a way that the feature factory is build at call and not init.

Furthermore, registering features now maintains a list of features, that is build in the build function. The docs need to be updated in an appropriate way, so that it is clear when available_features is updated.