Fix(env): a response window shows nothing until it closes

[Haor]

Two things let a player learn, from what the env offered them, what the other players could do or had done with the same discard.

1. A double ron was settled one winner at a time

In red_mahjong the first RON was paid, took the riichi sticks and went into action_history before the next player who could ron the same tile had answered. That player decided knowing whether the first had won, and for how much, ura dora included. On a 三家和 the third player saw two wins booked and could void both with a RON. This is the known limitation in the v0.1.6 release notes.

_ron now keeps each RON in pending_rewards and pending_winners, and _settle_ron pays them all, clears the sticks and records the RONs, in the order they were asked, on the step that closes the window: the last candidate's RON or PASS. A 三家和 has paid nothing, so there is nothing to roll back, and pending_kyotaku is removed. A window closed by a PASS keeps target and leaves the turn with the last winner, as one closed by a RON does, so the closed round does not show that someone declined.

2. The offered calls depended on the other players' hands

In both envs, _claims_open_to offered a player who could ron and pon (or chi) a discard the pon only when nobody else could ron (or pon) it, so the prompt told them about another player's hand. In a double ron it also told the second candidate what the first did: after a RON only the RON column was kept, after a PASS the whole row.

A player who can ron now always answers RON or PASS first; a declined ron leaves the pon or chi, offered once the higher claims are answered, as #87 already did when claims competed. A RON also leaves the rows in players.legal_action_mask exactly as a PASS does, so a candidate's own stored row does not show it either.

Behavior changes

• A player who can both ron and call a discard answers RON/PASS first, then the call.
• RON enters action_history, and ron payments enter score and rewards, only on the step that closes the window.
• In red_mahjong's EnvState, pending_kyotaku is removed and pending_winners is added. no_red_mahjong has no multiple ron and no such fields.
• A state serialized before this change does not load as is, and a caller that booked each RON's payment on its own step now gets the whole window's payment on the step that closes it.

Tests

• New tests/red_mahjong/test_ron_window.py, starting from real discards and an added kan:
  • a later candidate's observation, offered mask and stored rows are identical after every way the earlier candidates can answer (double ron, triple ron, robbing kan);
  • every RON/PASS combination pays exactly once, when the window closes;
  • three RONs abandon the round without paying anything;
  • the next round follows the closed window under auto and dummy_share;
  • a ron on a riichi declaration leaves the riichi unmade, and a declaration everyone passes stands at the next draw;
  • a declined ron still leaves the pon.
• no_red: a ron candidate is offered the same calls whoever else can claim the tile.
• Updated because they pinned the old behavior: the lower-call tests of both envs, and red's _claims_open_to, double-ron and triple-ron tests.
• 22 of the 24 new tests fail on main; the 2 that pass guard unchanged riichi behavior. Full suite: 334 passed, 2 skipped.

Tested on CPU with Python 3.12 and JAX 0.11.1, using the CI command (visualization tests excluded).