It calls attr.validate() (which calls the validators), and recomputes
the hash of HashableObject instances.

A future commit will also make it check the raw_manifest attribute
when relevant

It's just simpler this way

For the sake of completeness (a future commit may depend on it).

For now it is filled from 'offset' and 'negative_utc', but it will replace them
in a future commit.

This is to simplify and add support for more 'weird' offsets we do not currently
support.

This reverts commit c525484e.

It enables to easily check if a path exists from a root directory.

Since rDMOD8d96dfedee34203a4118e48a6208ee507511590b, directory entry
names are validated in DirectoryEntry model and thus must not contain
any slash characters.

So update directory_entries_d hypothesis strategy to ensure such names
are generated to avoid flaky tests.

Enable to compute md5 sum through the hashutil.MultiHash class.

Nevertheless, md5 is not put in DEFAULT_ALGORITHMS set and must
be explicitely requested by client code.

Related to T2400

I don't know any instance of these, but there is no harm in checking them.

1. hashes are now repr()ed as `hash_to_bytes("1234...")` instead of b"\x12\x34..."`
2. SWHID objects are now repr()ed as `CoreSWHID.from_string('swh:1:...:1234...')`
   instead of `CoreSWHID(scheme='swh', version='1', object_type=..., object_id=b'\x12\x34')`
3. enums are now repr()ed as `MyEnum.NAME` instead of "<MyEnum.NAME: 'value'>`

Thanks to these three changes, using repr() on a model object now prints
a string that can be pasted directly in a `.py` file to write a new
test case.

The previous replaced attrs-strict's type validator with our own,
stricter and faster, validator.

However, the strictness can be a burden in other packages;
for example, swh-storage tests rely on it to insert dummy data that raises
exception when accessed, and it would be hard to do while using the exact
expected type.

This commit reverts the strict behavior, but keeps the performance
optimization, by always checking with type equality, but in case type
equality fails (which would raise an error before this commit), it gives
the value a 'second chance', by trying isinstance.

This means that, outside tests, isinstance should not be used at all,
or very rarely.

This reimplements attrs_strict.type_validator(), using type equality
instead of isinstance.

This makes my checksum validation script (that mostly just instantiates
model objects, computes a checksum, then discard) run twice as fast.

For consistency, as the classes are now in swhids.py

+ raise warnings

1. Add a warning
2. Move identifier/manifest documentation to git_objects.py
3. Remove all imports of that module.

Motivation:

* SWHID classes were moved to swhids.py
* manifest computation functions were moved to git_objects.py
* Only reexports and trivial wrappers of model.py remain

They are not used anywhere.

Since they are used by the vault for non-identifier-related stuff,
I think it makes sense to move them to a new module.

identifiers.py is now an empty shell, as all its features were moved to
other modules and it only contains reexports and backward-compat
functions. Therefore, it should be considered deprecated from now on.

Refactor identifiers & model to make *_git_object() functions work on model classes instead of dicts

Since we now use these classes everywhere, computing hashes required
using to_dict() just to compute identifiers, which can be a performance
bottleneck in code computing many checksums.

A future commit will make identifier computation use the attrs classes,
which are strict about what they accept.

identifiers.py initially worked only on bare sha1_git.

I chose to add the SWHID classes in that module because
of the name, but the SWHID code didn't actually interact
with the other functions in the module, so it now feels
out of place to me.

We're about to have a Bazaar loader

This class is a mixin that only works with classes that define
this attribute, so it makes sense to declare it.

It also allows generic functions (that take a HashableObject
parameter) to access it without 'type: ignore'.

This allows distinguishing multiple potential versions of the mapping
between external objects and their counterparts archived in Software
Heritage, for instance when a loader has a backwards-incompatible change
that should result in objects being loaded again.

The field defaults to zero, in which case it's backwards-compatible with
the previous implementation in terms of identifier computation.

We agreed a while ago they should be IRIs and not just URIs.

This will trigger crashes in swh.storage.cassandra, as
currently expects (wrongly) that origin urls are ASCII.

* empty fetcher name or version is not accepted by cassandra
  (and is nonsensical anyway)
* ditto for non-ASCII (and any non-printable is nonsensical)
* null bytes/chars are accepted by neither postgresql or cassandra

This is the only value we should use from now on;
and the value is ignored by swh-storage anyway.

It will be used by swh-web.

Closes T3393

- add typing annotation to avoid such error in the future

Fixes T3383

the problem was for datetime<epoch, the timestamp is negative, but since
it's a float that includes the microseconds, if both are true (< epoch
and microsecond > 0), then the computed (int) timestamp was off by one.

Add dedicated tests for this.