[AnimalShogi] precompute the CAN_MOVE table once instead of per _can_move call

[gweber]

What

_can_move rebuilt the full (5, 12, 12) CAN_MOVE table — via a triple-nested vmap — on every call, only to index a single entry:

def _can_move(piece, from_, to):
    def can_move(piece, from_, to): ...        # pure movement geometry
    CAN_MOVE = jax.vmap(jax.vmap(jax.vmap(can_move, ...)))(jnp.arange(5), jnp.arange(12), jnp.arange(12))
    return CAN_MOVE[piece, from_, to]

The table depends only on movement geometry, so this lifts it to a module-level constant (as the Zobrist tables already are). _can_move runs ~1700× per _legal_action_mask (132 move checks + 132×12 in _is_checked), so the rebuild dominated.

Behaviour & perf

Pure refactor, no behavioural change — legal_action_mask matches the previous implementation across 250 states, and tests/test_animal_shogi.py passes (5/5). Measured ~1.57× faster _legal_action_mask (1.42 → 0.91 ms, B=200, CPU) and ~1.2× faster compile. (XLA does not fully constant-fold the per-call rebuild, so this is a real runtime win, not only a tracing/clarity one.)