create a downsampled iterator so that downsampling becomes less dependent on clone

[ctb]

per comment on #3342, #3342 (comment)

@luizirber speaketh:

I wonder if we can do (in a future PR, not this one) a new .downsampled_iter(scaled) for operations like count_common, and avoid the conversion.

The downsampled iter would iterate over values that are in the appropriate scaled value, but wouldn't need to create new minhash sketches (can reuse the largest one and stop returning values once they go over max_hash, for example)

[ctb]

maybe explored here? #3394

[luizirber]

maybe explored here? #3394

In a similar direction, but not quite. The downsample iter is more similar to .iter_mins() or .iter_abunds() in #3394, but there I'm only using iterators directly to calculate the intersection.