Unify caching mechanisms and separate their management from Interpretations

[eb8680]

Addresses #633

This PR implements a new weakref-based caching system that unifies and generalizes those introduced in #726 and #594

Specifically, it adds a new context manager effectful.internals.runtime.cache such that within a cache()-delimited block, repeated calls to evaluate with the same expression or data structure and active interpretation are cached.

This new mechanism is backed by a new standalone utility module effectful.internals.weak, based on a similar utility module torch.utils.weak in PyTorch, which implements a couple of different cache data structures generalizing weakref.WeakKeyDictionary:

• WeakIdKeyDictionary behaves like WeakKeyDictionary (including forbidding keys which cannot be weakref.refed, like tuples and ints) but uses object identity rather than object equality/hashing to discriminate between keys. (This is the main thing provided by torch.utils.weak)
• IdKeyDictionary, which stores strong references to keys that it discriminates between using identity
• AutoIdKeyDictionary, which extends WeakIdKeyDictionary and IdKeyDictionary's identity-based semantics by attempting to use a weak reference for a new key and falling back to a strong reference when that is disallowed by the key type, and ensuring that weak reference and strong reference keys cannot conflict with one another.
• weak_memoize, a helper function that can use these data structures to memoize a single-argument function analogous to functools.cache

The PR also includes a number of small caching-related semantics-preserving performance fixes, and a benchmark derived from a RoboTL issue illustrating both asymptotic and constant-factor improvements on a representative workload.

[eb8680]

Splitting into two separate PRs for easier review