Invalidate the presorted hint when clearing Stats caches

[ac1982]

The change

Stats.clear_cache() now discards the is_sorted hint as well as cached measures. Its documented use after modifying Stats.data could otherwise recompute incorrect values from an order that no longer holds:

stats = Stats([1, 2, 3], is_sorted=True)
stats.data.append(0)
stats.clear_cache()
(stats.min, stats.max, stats.median)  # before: (1, 0, 2.5); expected: (0, 3, 1.5)

Reset the hint when clearing the cache, document that behavior, and cover repeated mutations with both copy modes. The is_sorted=True optimization remains available until the caller invalidates the cache.

The regression failed on master. After the fix, python -m pytest --doctest-modules boltons tests -q passed: 679 tests and 12 subtests on Python 3.12/Linux. Other platforms were not run locally.

Related #534 preserves mode tie-breaking while computing sorted measures. That PR leaves _is_sorted set after clear_cache() too, so it does not fix this presorted-input mutation case. This change does not alter mode tie-breaking or introduce a sorted-data cache. Open/closed searches for clear_cache, is_sorted, and Stats cache found no matching fix.

The backstory

Same contribution experiment and tooling as #569. While reading Stats' explicit mutation/cache contract, Codex reproduced wrong minimum, maximum, and median results after following the documented clear_cache() advice. This is a locally constructed regression, not a production incident.

AI assistance

• Harness/tooling: OpenAI Codex, shell, pytest, GitHub CLI.
• Model: GPT-6-based Codex.
• Original prompt: “OK，继续今天的工作吧，magpie 只算做 addtional 任务，不记作当前天的 30，所以今天总共任务是 32 个。” Earlier instructions require substantive bug reproduction, rules/duplicate checks, minimal fixes and tests without reducing quality for quantity. Codex selected this independent cache-invalidation defect during repository inspection; no Stats-specific prompt was supplied.

[Wkhanex]

Appending out-of-order values and calling clear_cache() now agrees with Python's statistics results. I checked successive appends with both copy modes, including a new negative minimum: median, mean, minimum and maximum are all recomputed.

Checked cb330b389ef5 with standalone Python 3.12.14 checks (24 passing assertions). I haven't run the full upstream suite.