consider further optimization/exploitation of FrozenMinHash

[ctb]

Over in #2128, I wrote

Also while I am at it, any speed/memory advantage using FrozenHashes?

Not a big one. The main goal of FrozenMinHash is to enable future optimizations where we don't make new copies of FrozenMinHash objects when we don't need to. It also helps deal with object ownership issues that complicated the Rust-and-Python layers. You can read more #1616 if you're really interested ;).

...and then when debugging #2132, I immediately found a bunch of situations where we could cache various computations done in FrozenMinHash such as __len__.

So.

I think we could do some judicious profiling to guide the addition of cached properties to the FrozenMinHash class based on immutability of the objects.

[ctb]

(being explored in #2135)