perf(sichuan): reuse ordinal hand counts

[TypeThe0ry]

What changed

• Replaces tile-name parsing with MahjongTile.ordinal() lookup tables.
• Builds one 27-entry count array and increments/decrements candidates in place.
• Uses count-based recursion for waiting-tile checks.
• Retains the existing complete shape construction, ordering, fixed-meld handling, and red-five normalization.

Why

The Sichuan evaluator repeatedly parsed tile names and rebuilt collections for every waiting-tile candidate.

Impact

Waiting-tile order and complete winning shapes remain byte-for-byte compatible in differential tests. No public API or rules behavior changes.

Evidence

Paired JMH median reductions:

• canWin: about 48.9% time and 76.8% allocation
• evaluate: about 52.4% time and 56.1% allocation
• waitingTiles: about 78.3% time and 97.2% allocation

Validation

• SichuanHuEvaluatorTest, including randomized legal hands and fixed melds
• spotlessCheck
• full integration test suite passed before commit splitting

[Arbousier1]

Formal trusted A/B run did not pass: INCONCLUSIVE_NOISE_OR_NO_GAIN. Evidence: https://github.com/EllanServer/MahjongEngine/actions/runs/29645450524. Closing per the performance gate policy.