This saves 30KB in the final archive.

Needed for debian builds to pass.

The interfaces were a bit messy, as bundle types sometimes included the object
type ('directory' was also a bundle type) and sometimes did not.

Instead, the bundle types now do not include the object type at all, which
means users do no have to care about what cooker will cook their bundle,
they just need to provide a bundle type.

Instead, the information about the object type is now carried along the
object id, in a SWHID.

This allows simplifying interfaces and removing multiple convertions
(especially between hex and bytes).

This is unfortunately a pretty large commit, as it is easier to change
everything at once, than to try to make it incrementally and adding
temporary convertions at some moving borders.

It also invalidates the existing cache entirely, as it used bundle types
and object ids as keys, which are now changed.

Using this type narrowing trick:
https://mypy.readthedocs.io/en/latest/literal_types.html#exhaustive-checks

Instead of parsing self.bundle_type on the fly in multiple places.

'obj_type' was confusing, because it means 'the type of objects in the
vault backend', which is actually made of an object type (dir/rev/...)
AND the cooker name.

Plus, I want to reuse the name 'obj_type' in a future commit to reference
what is actually the object type, in the swh-model definition.

It allows me to tweak the code to experiment

This should be a considerable performance improvement, as we
don't need to query swh-graph for every head (which includes
a lot of duplicates).

There is nothing interesting about this traceback because it only
shows frames in the objstorage and the CLI.
The actual error is always printed before with its own traceback
before this one is shown.

So the 'load' and 'check' parts of the tests can be reused by future
snapshot tests.

So its 'load' and 'check' parts can be reused for testing snapshots

which happens in the Debian builder environment

It expects the path to be a string rather than a pathlib object.

Dulwich 0.19 is in Debian buster so we need to support it.

This decreases warnings during tests.

Git 2.21 changes the error message raised by git-fsck when an object
hash does not match the content:
https://git.kernel.org/pub/scm/git/git.git/commit/?id=01f8d5948a7f370c42d9fe2deb724139a1bfcb7b

Those were missed in the first part of the implementation. They got revealed by the
debian build.

Related to T3412

It's not packaged nor ready yet as some internal reworking has been discussed.

Related to T3412

They used to crash, now they are replaced with a placeholder,
like the gitfaster cooker does.

instead of going through swh-storage.

This also allows batching queries, so it should be more efficient overall.

It is considerably faster (30% less run time on an average repo)

As it does not need to join with the content table.

On small repositories with a warm cache, it doesn't seem to matter much, though.
But it's also closer to a feature swh-graph will provide in the future,
so it's a win anyway.

This will allow batching large groups of objects, instead of being limited
to those given as argument from a parent.