Add four-player Riichi Mahjong (single hand)

[jaisinha77777]

Per #843 (see comment there flagging this before opening the PR, per #1511's guidance for larger contributions).

Adds python_mahjong_riichi: one hand of four-player Japanese Riichi Mahjong, Python-only.

The calling problem. Pon/Chi/Kan/Ron is a real-time interrupt mechanic with no direct mapping onto OpenSpiel's simultaneous-move or turn-based dynamics — there's an old, unresolved thread on exactly this at #979. This implements it as sequential, priority-ordered polling after each discard: every Ron-eligible player is asked first, in turn order (multiple simultaneous ron is supported); only if nobody rons is the single Pon/Kan-eligible player (if any) asked; only if nobody calls that is the single Chi-eligible player (the discarder's next player) asked. This reproduces standard priority (Ron > Pon/Kan > Chi) without requiring true real-time simultaneity.

Scoring (mahjong_riichi_utils.py, kept separate from game flow so it's independently unit-testable): standard hand-shape decomposition (4 sets + pair, chiitoitsu, kokushi), the standard ~25 named yaku, fu calculation (wait type, open/closed triplet-vs-kan distinction, shanpon-on-ron vs shanpon-on-tsumo, double-wind pairs), dora/aka-dora/ura-dora, and the mangan-through-yakuman point/payment tables. The common yakuman are covered (kokushi, suuankou, daisangen, shou/dai-suushii, tsuuiisou, chinroutou, ryuuiisou, suukantsu, chuuren poutou, tenhou/chiihou); a handful of rare local-rule yaku (e.g. renhou) are intentionally not included.

Also implemented: riichi declaration with furiten tracking (both permanent, from a player's own discards, and temporary, from declining a legal ron), chankan (robbing a kan), and the abortive draws (kyuushu kyuuhai, four-kan, four-riichi, four-wind-discard).

Scope: one hand per episode (deal through win or exhaustive draw), not a full multi-hand match with dealer rotation and score carryover across hands — same pattern as e.g. leduc_poker modeling one hand rather than a session. returns() is the point swing for the hand.

Testing:

• mahjong_riichi_utils_test.py: 27 tests on hand-shape decomposition, fu/yaku detection (including a regression test for a sanshoku-doujun bug caught along the way — it originally only matched runs starting at "1"), dora counting, and the payment tables, checked against hand-derived reference values.
• mahjong_riichi_test.py: 9 tests covering the deal structure, action-encoding consistency, a directed Tenhou-tsumo scenario, a directed ron + furiten scenario (confirming furiten blocks ron across a player's entire wait-set, not just the declined tile), riichi discard-locking, pyspiel.random_sim_test with serialization, and a zero-sum-returns check across many full random games.
• Playthrough generated and verified to round-trip via generate_playthrough.replay.
• Registered in open_spiel/python/games/__init__.py, CMakeLists.txt, and pyspiel_test.py's expected-games list; added a row to games.md.

Happy to iterate on the calling-priority design in particular, since that's the part without existing precedent in the codebase.