Fix MJAI replay start_kyoku wall initialization

[hong0311]

Summary

Fix an off-by-one bug in MJAI replay initialization.

apply_mjai_event(start_kyoku) was leaving replay state in the post-dealer-draw position, even though MJAI logs emit the dealer's first draw as a separate tsumo event. When that logged tsumo was applied, the wall became one tile too short for the rest of the round.

This mostly goes unnoticed, but it breaks legality checks near the end of the hand where rules depend on the remaining wall count. In the reproduced case, a real chi window was reconstructed as Ron/Pass, and select_action_from_mjai(...) failed to resolve the source event.

What changed

• Reset 4-player MJAI replay state to the correct pre-dealer-draw wall length in start_kyoku
• Apply the same fix to 3-player replay state
• Reset round-scoped replay fields during start_kyoku so replay starts from a clean round state
• Add regression tests covering the pre-tsumo wall length in both 4p and 3p

Why this fixes the bug

For MJAI replay, the correct sequence is:

• start_kyoku -> wall is still pre-dealer-draw
• first tsumo -> consume exactly one tile

Before this patch, replay effectively started one step too late, so the first logged tsumo consumed an extra tile and shifted the rest of the round out of sync.

Repro

Using 2024010109gm-00a9-0000-1fe4306c.json:

• before this patch, after player 2 discarded 9s, seat 3 saw Ron/Pass
• after this patch, seat 3 correctly sees Chi/Pass
• select_action_from_mjai({"type":"chi","actor":3,"target":2,"pai":"9s","consumed":["7s","8s"]}) now resolves successfully

Testing

• cargo test -q apply_mjai_event_start_kyoku -- --nocapture
• cargo test -q test_apply_mjai_event_honor_and_red_tiles -- --nocapture

Manual verification:

• converted the repro paipu to MJAI events
• replayed the failing round via apply_mjai_event(...)
• confirmed the response window now matches the source chi

[smly]

@hong0311 Thanks for the PR! This looks like a solid and important bug fix for the MJAI replay. Really appreciate it! I'll verify the repro using the paifu mentioned in the description just to be safe.

P.S. Now I remember — I originally implemented replay for mjsoul logs first, and added MJAI log replay later. In mjsoul logs, the dealer receives 14 tiles in the start_kyoku equivalent. That's exactly where this off-by-one came from...

[smly]

LGTM!

[smly]

Just to note: the minimal fix for the off-by-one is self.wall.tiles = vec![0; 136 - 13 * 4] (and the corresponding line for 3p). The rest are preventive resets and regression tests, which are welcome.