Add functools.cached_property backport

[maffoo]

The @functools.cached_property decorator is really useful but not currently available in cirq since it was introduced in python 3.8 but we still support python 3.7. This adds backports.cached_property to make cached properties available in python 3.7, using import logic in cirq._compat to get it from the standard library or backport module, as needed.

[MichaelBroughton]

Does doing something like:

@property
@functools.lru_cache(maxsize=None)
def thing(x):
   stuff(x)

not work ?

[maffoo]

@MichaelBroughton, no, that doesn't work generally because it it requires all parameters to be hashable, including self. It also attaches the cache to the class, so that the cache will grow without bound if you create lots of instances. With cached_property the value is cached on the instance instead, so values get gc'd with the instances they are attached to.

[MichaelBroughton]

Fair enough. Last question: Why should it be Cirq's responsibility to provide this wrapper for cached_property ? Do we have use cases in the actual code base ? I don't think we should expose cirq.cached_property just as an unrelated convenience method for users otherwise.

[maffoo]

Agree we shouldn't export this at the top level, that's why I put it in cirq._compat but did not export outside of that. There are quite a few places where we implement ad hoc property caching that I think could be converted to use this, but it's more of an internal implementation convenience than something we want to expose to users.