Use metaclass to subclass errors to allow better pickling

[BramVanroy]

When trying to pickle with dill, as happens under the hood when processing a HuggingFace dataset, I found that spaCy couldn't be pickled (related issue: huggingface/datasets#3178). I found that the cause was dill.Pickler that recursively pickled all objects (Pickler(file, recurse=True).dump(obj)). This is preferred in datasets: such iterative pickling allows for fingerprinting, i.e. keeping track of all the processing done to a dataset, deterministically. If the fingerprint is the same as an earlier processing attempt, then a cache can be used and the processing (e.g. tokenizing with spaCy) does not have to be done again. Very useful and great to save time and processing.

The culprit in the spaCy codebase that could not be pickled in this way, was this part

spaCy/spacy/errors.py

Lines 4 to 15 in f1bc655

	def add_codes(err_cls):
	"""Add error codes to string messages via class attribute names."""
	class ErrorsWithCodes(err_cls):
	def __getattribute__(self, code):
	msg = super(ErrorsWithCodes, self).__getattribute__(code)
	if code.startswith("__"): # python system attributes like __class__
	return msg
	else:
	return "[{code}] {msg}".format(code=code, msg=msg)
	return ErrorsWithCodes()

The class ErrorsWithCodes cannot be pickled here because it does not have a fixed signature in the global scope (its superclass is not known beforehand). After digging through the Internet (thank you Stack Overflow!) I found that a metaclass is the way to go, as implemented in this PR. In my opinion, this is also more structurally clear than before. Regular subclassing a metaclass instead of dynamically created a class based on a subclass with a decorator. The metaclass defines the magic methods that are required to access the class' attributes, rather than accessing instance attributes.

I modified the error test accordingly, which completes successfully.

After this modification, recursive dill pickling works flawlessly as well.

closes #9584

Types of change

Enhancement

Checklist

• I confirm that I have the right to submit this contribution under the project's MIT license.
• I ran the tests, and all new and existing tests passed.
• My changes don't require a change to the documentation, or if they do, I've added all required information.

[honnibal]

Thanks, the change makes sense to me, and it's nice to support pickling with dill.