v0.3.0

On skewed streams the delete branch churns on the same hot keys, and V8
keeps deleted entries in the Set's bucket chains until the table is
rebuilt, so lookups walk ever-longer chains. This copies the set once
holes pile up past max(|X|, 1024): membership, order and randomness are
untouched, so same-seed results are bit-identical to 0.2.0.

The bench gains a skewed scenario and a below-threshold one, and the
README a "stream shape" table that also says where the estimator buys
nothing over a plain Set. npm run bench reproduces the numbers.