Add random directory sampling policy

This makes use of the discovery algorithm introduced in
``swh-model``, which should help speed up large (think Linux
kernel or way larger) scans.

Most of the time is spent walking the on-disk directory and hashing,
which is where the new optimizations in ``swh-model==6.5.0`` should come
in handy. Python is close to its limit in that regard, some future
endeavor should look into setting up SWH for native extensions.