feat(ruleset_strategy): add round_score "rank" (per-episode placement scoring)

[daveey]

What

Adds an opt-in rank-by-episode round scoring mode to the ruleset_strategy commissioner. Requested for the agricogla league (the metta-side config + image bump is a follow-up).

scoring.round_score: "rank"

ScoringConfig.round_score previously only accepted "mean" (round score = mean of a policy's per-episode scores). Adds "rank":

• Within each episode, policies are ranked by score and earn N..1 rank points — the winner of an N-policy episode gets N, last gets 1, ties share the better placement (computed as N - (#strictly-higher)).
• A policy's round score is the mean of its per-episode rank points across the episodes it played. Margins of victory are discarded — only placement each game matters.

How

• complete_round now delegates per-policy round scoring to an overridable _round_scores_by_policy(entries, episode_results) -> (scores, ranked_counts). The base BaselineCommissioner keeps mean scoring; RulesetStrategyCommissioner switches to rank points when scoring.round_score == "rank". The ranking/metadata assembly is unchanged.
• Rank rounds are tagged with a distinct score_kind (rank_episode_round_score) via RankingConfig.result_metadata/filter_metadata, so switching a league from mean → rank filters the now-incomparable prior-regime round results off the commissioner leaderboard instead of blending two score scales.
• scoring_mechanics describes the rank scheme for the division description.

Default stays "mean", so every existing config behaves exactly as before.

Tests

• New test_ruleset_strategy_rank_round_score_uses_per_episode_placement: same episode inputs as the existing mean test, asserts round scores become the mean per-episode rank points and the rank score_kind tag is applied.
• Full suite: 91 passing. Changed files are ruff-clean (two pre-existing unused-import warnings in utils.py are untouched).

Follow-up (not in this PR)

• metta: set agricogla's agricogla-commissioner.yaml to scoring: {round_score: rank} and bump the commissioners-default image digest once this is merged + the image is rebuilt/published.
• The app-backend leaderboard needs no change — it aggregates the commissioner's round score, which becomes mean per-episode rank points.

🤖 Generated with Claude Code

[nishu-builder]

@KyleHerndon note that right now, with division leaderboard computation and commissioner leaderboard computation split the way we do, and with our UI only reflecting commissioner-reported description, we force the commissioner to abstraction-leak by describing how its roundresults get managed by app-backend

[nishu-builder]

sorry more plainly: ideally we wouldnt need commissioners to say that their roundresults get 2h-ewma'd; they shouldn't need to know about or speak about it, and can't enforce it