Follows swh.model's evolution for the Revision model class.

In Postgresql, store the extra_headers as a bytea[][].
Ensure data present in postgres with extra_headers in the metadata field
are properly supported by the pg-backed storage.
Get rid of the (now useless) git_headers_to_db() converter function.

In Cassandra, store them as frozen<list<list<blob>>>.

This endpoint is not really useful since the origin_add() can be used
instead.

Using a single API endpoint would also make the API a bit more consistant
(most other endpoints only provide a xxx_add endpoint) ; having a single
endpoint per object_type make is enough and make the whole API simpler.

The CLI command is included as well as `swh storage replay`.

Copied test test_replay, in test_cli.py, should be identical.

This componant makes more sense in the swh-storage package.

Which is a maintained fork of retry

- use pytest instead of unittest.TestCase plumbing
- extract data from the TestStorageData into a data `storage_data` module;
  this module also provide a simple helper `StorageData` class that mimics
  the original class (access by attributes),
- implement a series of pytest fixtures for these storage specific tests,
- get rid of most hypothesis-based tests,
- replace usage of the use_url hypothesis boolean statetgy by
  pytest.mark.parametrize fixtures; this allows to prevent from
  the need of resetting the storage, since tests are truly executed
  twice (thus with a new swh_storage),
- refactor test_db to use pytest-postgresql.

Disable (xfail) tests from test_snapshot.py, test_api_client
and test_in_memory for now.

Fixes/refactorings come with following revisions.

in particular, ignore django.* (only indirectly shipped by hypothesis) and
allow unused type: ignore (on swh.journal)