Make props related functionality a part of the Op baseclass

[ricardoV94]

Description

This is in MetaType, but AFAICT only ops ever have props

pytensor/pytensor/graph/utils.py

Lines 193 to 234 in 17748b7

	props = dct.get("__props__", None)
	if props is not None:
	if not isinstance(props, tuple):
	raise TypeError("__props__ has to be a tuple")
	if not all(isinstance(p, str) for p in props):
	raise TypeError("elements of __props__ have to be strings")
	def _props(self):
	"""
	Tuple of properties of all attributes
	"""
	return tuple(getattr(self, a) for a in props)
	dct["_props"] = _props
	def _props_dict(self):
	"""This return a dict of all ``__props__`` key-> value.
	This is useful in optimization to swap op that should have the
	same props. This help detect error that the new op have at
	least all the original props.
	"""
	return {a: getattr(self, a) for a in props}
	dct["_props_dict"] = _props_dict
	if "__hash__" not in dct:
	def __hash__(self):
	return hash((type(self), tuple(getattr(self, a) for a in props)))
	dct["__hash__"] = __hash__
	if "__eq__" not in dct:
	def __eq__(self, other):
	return type(self) is type(other) and tuple(
	getattr(self, a) for a in props
	) == tuple(getattr(other, a) for a in props)
	dct["__eq__"] = __eq__

Generating them dynamically like this means linters / mypy don't believe in op._props() or op._props_dict(). I don't know why this has to be implemented as methods in the MetaClass instead of just in the Op baseclass. Would be great if someone gave a try at it. It also means we could probably get rid of the whole MetaType class, since the only other thing is a string convenience, which definitely need not be defined there.

May also be worth investigating if we can use frozen dataclasses for Ops and automate the props logic (see aesara discussion, but that needs not be done together with addressing the main issue here.

[PabloRoque]

@ricardoV94 I am giving it a go in this PR inside my fork [so I don't make noise around here].

Some learnings:

• Type also inherits from MetaType through MetaObject. This class acts as super of a bunch of other types.
  • We would then need to add props also in this class, not only in Op
• Some tests [e.g. CustomOpNoProps] would need to be hacked to retain previous behavior. This is a bit convoluted, because all Ops should have __props__.
• We would have to patch some MyOp definitions inside tests like this one to explicitly include __props__

[ricardoV94]

I don't think we need props in those other classes, let me know if they are actually used anywhere.

Can __props__ be None by default and make do work as before? If None we hash and match on identity as default Python does.