This ensures all instanciated model entities have valid types for attributes.

Related to T2308.

(pairing with @vlorentz)

Related to T2310

it was still here for bw-compat but should not be necessary any more.

this function does not need to be a local function of the to_dict namespace.

This will allow deduplicating code across loaders.

This lets us generate Content objects directly from a bytestring, with the
proper set of hashes auto-generated from the contents.

Required by loaders, when they can't parse the fullname.

Will be used by loaders.

They will be used by loaders, so they can deal only with
model objects, instead of having to do the same conversion themselves.

This removes the `data` and `save_path` arguments of `from_file` and
`from_disk`, as data loading is always deferred from now on.
To access it, users are now expected to either open the data files
themselves, or us `.to_model().with_data()`.

Can be useful to deduplicate code in swh-storage.

It already is in practice.

The current postgresql model refuses NULL values.

Related to P589

Related to T2188

Add support to automatically compute identifier in the following object models:
Directory, Release, Revision, Snapshot.

If the identifier is not provided as parameter, it will be computed when the model
is initialized.

we do not really need them to be mutable, plus we gain their instances now
being hashable, so we can add them in set() for example.

Instead of relying on attr.asdict recursion, we do recursion ourselves. This
simplifies a lot of the inherited to_dict() methods.

Needed by the replayer in swh-journal.

It kind of matches the current state of the postgresql storage, which
does not support it.

But keep support for deserializing from str, like
swh-storage does.

Code manipulating a Content object may want to access a hash
of configurable name; this method allows it to do that
without using getattr directly.

It denotes files whose length is unknown.

Loaders use it, but it is ignored by the pg storage.
However, as the Cassandra storage uses swh-model to validate
its input, it refuses this input from the loaders (and journal).

There is more repetition, but it's easier to read and
'%timeit Revision.from_dict(d)' is 5 times faster.