List extractor capabilities, e.g. in datalad wtf-output

[christian-monch]

This is a follow up of PR #351

In addition to the "interface-generation" that is made explicit in extractors with PR #351, we should make the capabilities explicit. More precisely, extractors should be able to state whether they support file-level, dataset-level, or file-level and dataset-level extraction.

As outlined here, we can add a class property named __extraction_modes__, that would be a non-empty subset of {'file', 'dataset'}. This could be, for example, reported in datalad wtf.