[Chess] Fix stale INIT_ZOBRIST_HASH: derive it from the Zobrist tables at import time

[gweber]

Bug

The Zobrist tables in pgx/_src/games/chess.py are generated at import time from jax.random, whose outputs are not stable across JAX versions (e.g. the threefry_partitionable default change). The hardcoded constant

INIT_ZOBRIST_HASH = jnp.uint32([1455170221, 1478960862])

no longer matches _zobrist_hash(GameState()) under current JAX (0.10.1 gives [1875025768, 182641400]). Since GameState() seeds hash_history[0] with the stale constant while every later entry is recomputed, the initial position's history entry never matches:

• Repetitions involving the start position are counted one short. Repro: the knight shuffle g1f3 g8f6 f3g1 f6g8 g1f3 g8f6 f3g1 f6g8 is a threefold repetition at ply 8, but pgx terminates only at ply 9 (when the position after 1.Nf3 repeats a third time instead).
• The repetition observation planes (rep0/rep1) are wrong for the startpos history entry.

import jax.numpy as jnp
import pgx._src.games.chess as C
print(C.INIT_ZOBRIST_HASH)            # [1455170221 1478960862]
print(C._zobrist_hash(C.GameState())) # [1875025768  182641400]  <- mismatch

Fix

Compute the constant from the actual tables at import time (same XOR-reduce idiom as _zobrist_hash), so it can never go stale again, under any JAX version. On JAX versions where the old constant happened to match, this is a no-op.

Note: gardner_chess.py and animal_shogi.py hardcode similar init-hash constants and likely have the same issue; this PR fixes chess only.