`Implements` should cache its calls (like `Provides` does)

[d-maurer]

plone/Products.CMFPlone#3829 describes a case where the adapter factory lookup cache of zope.interface grows without limit causing severe memory problems.
I found the cause: plone.dexterity uses Implements (instead of Provides) in its implementation of the providedBy logic. While Provides supports efficient caching, Implements does not as demonstrated by

>>> from zope.interface import Interface
>>> from zope.interface.declarations import Implements, Provides
>>> class I(Interface): pass
... 
>>> I1 = Implements(I); I2 = Implements(I)
>>> P1 = Provides(I); P2 = Provides(I)
>>> I1 == I2
False
>>> P1 == P2
True

Provides achieves this by caching its calls (in a WeakValueDict); Implements does not use such a cache.

While I think that plone.dexterity uses the wrong declaration type (Implements instead of Provides), I believe that Implements should cache its calls like Provides does to improve its caching support.

[jensens]

FTR, Plone's problem was fixed Aug 2023, in PR plone/plone.dexterity#187 (released as 3.0.3): Changed from Implements(*dynamically_provided) to Provides(cls, *dynamically_provided) in FTIAwareSpecification.__get__().