Reported issues with the registry procedure

[tlienart]

Below are some issues reported by a user "Y" trying to fully test that his interface to MLJ worked by setting up a local registry and checking that he could do "@load". I indicated that, for testing purposes, it was better to just go with the MLJBase API directly but the issues are still interesting for possible future improvements (PS: given that these issues are not mine, I may not be able to fully comment on them)

1. trying to register an abstract model type as opposed to a full model type will fail in an obscure way --> cleaner error messages would be good
2. the registry manifest h as examples of repeated symbols for the implemented_fields field
3. maybe write explicitly how to fully test the model and that provided they can test it with just MLJBase it will necessarily work with @load or explain how they could locally test the @load.