Refactor apriori algorithm - #14579

@cclauss I dug into this one. The algorithmic restructuring is genuinely useful — it replaces the old ad-hoc prune() (which counted item tuples and never did true subset pruning) with the textbook Apriori shape: generate_candidates doing the F(k-1) ⋈ F(k-1) join, has_infrequent_subset doing proper (k-1)-subset pruning, and support counted with frozenset.issubset. That's a real correctness/readability win over main.

But it can't merge as-is — two blockers:

1. All doctests were removed. apriori() and the deleted prune() lost their >>> examples; doctest.testmod() now tests nothing. This repo requires doctests (or a dedicated test) for every function, so we'd be reducing coverage. The refactor needs fresh doctests that pin the new output.

2. The return type is inconsistent (and the annotation is wrong). The signature says -> list[tuple[frozenset, int]], but 1-itemsets are appended as (next(iter(i)), c) → a str first element, while k≥2 itemsets are (sorted(c), count) → a list first element. Running it on load_data() with min_support=2 gives:

   ('milk', 4)          # str
   ('bread', 2)         # str
   (['bread', 'milk'], 2)  # list
   

   Callers can't consume that uniformly, and it doesn't match the annotation. Also note the output contract changed vs main (1-itemsets are now included), which should be called out in the PR description.

Suggested path to merge: make the 1-itemset rows (sorted(i), c) so every row is (list[str], int), correct the annotation to list[tuple[list[str], int]], and add doctests for both helpers and apriori(). If the author does that, I'd be happy to re-review — the underlying refactor is worth landing.