We do not know the encoding beforehand so manipulate every blob coming
from the archive as bytes.

- Change class interface to match async interface (eg: iter -> aiter)
- Sync code is now running with `run_in_executor`
- Use aiosqlite as the async interface for sqlite

This is necessary to read very long files, or use a terminal pager.

See D4139.

Better class abstraction. Now the information is less scattered around
the code (especially for fuse.py). Each FUSE entry virtual type is
self-contained and defines entirely how it is used, listed, read, etc.

- Create a pytest fixture to run swh-fuse in a separate process and
  setup an offline mock web API
- Add basic cnt/dir tests

More flexible and easier to tweak settings. For example we can now have,
the metadata cache in-memory while the blob cache is on-disk:

  cache:
    metadata:
      in-memory: true
    blob:
      path: /cache/dir/blob.sqlite

Rework the cache architecture. The cache code is now separated into
multiple classes (one for each cache type, and one global). Every cache
is now used as a context manager and must implement `__getitem__` and
`__setitem__` to access/update the underlying cache information.

To stop the process, run: `fusermount -u MNTDIR`.

- Variables/comments rewording
- Use SQLite context manager instead of manual commit
- Get a list of swhids as input in the CLI

FUSE bindings to locally mount and navigate (as a virtual file system)
the Software Heritage archive.

- Basic CLI (usage: swh fuse mount SWHID PATH)
- Handle `cnt` and `dir` artifacts
- User-level commands working: ls, cd, cat
- Metadata and blob are cached using SQLite

in particular:
- use docs/README.rst to reuse it via include
- add docs/cli.rst documentation for CLI