2nd part of #813, after PR #863

This PR implements the core functionality needed for experimenting with "flexible fusion", i.e. making it possible to apply Annif backends not just on document text, but metadata fields. It enables the following:

• The CSV short text format can now contain extra columns/fields for metadata. Internally, they are stored in the new metadata field within Document objects.
• A new transform select can be used to select which parts of a document (main text and/or metadata fields) are given as input to the backend. For example select(title) or select(title,description,text)

Implementation notes

• I had to change the namedtuple Document into a real class in order to better support optional fields. Now only text is a mandatory field, while subject_set and metadata are optional (defaulting to empty set and empty dict, respectively)
• I extended the transform functionality. Previously transforms could only operate on text. Now transforms can implement either transform_text (to transform text only, similar as before) or transform_doc (transform whole Document objects). The latter was needed to implement the select transform.
• I had to change all the various suggest methods on both projects and backends to accept a Document (or batch of Document objects) instead of text. This lead to a lot of churn, but it's all in a single commit b0bb163.
• Currently, this PR is targeting the PR branch of Flexible fusion part 1: CSV short-text document corpus format #863 because this is built on top of that. It needs to be merged after Flexible fusion part 1: CSV short-text document corpus format #863.

What is still missing

This PR implements only the minimum needed for experimentation with custom metadata and select transforms: it should now be possible to train and evaluate models with custom document metadata, as long as the CSV short text corpus format is used.

These parts are still missing, to be implemented in future PRs:

• support for metadata in the full text corpus format (for example via YAML front matter, or a separate file for metadata)
• CLI suggest command should allow specifying extra metadata fields, probably as command line arguments
• REST API suggest method should allow specifying extra metadata fields in the request + the http backend should support it
• Web UI should allow entering metadata fields