chore: silence flax/optax deprecation warnings in trainer + tests

[NorbertRop]

Summary

Bundled cleanup of the deprecation warnings CI was flagging on every run. Three migrations, all mechanical, no behavior change.

What	Where
`optax.global_norm` → `optax.tree.norm`	`src/jaxpot/rl/trainer.py` (2 sites)
`nnx.State.merge` → `nnx.merge_state`	`src/jaxpot/rl/trainer.py`, `tests/test_checkpoint_roundtrip.py`
`variable.value` → `variable[...]` (Array) / `.get_value()` / `.set_value()` (struct)	`src/jaxpot/rl/{trainer,ppo_trainer,alphazero_trainer}.py` (~9 sites)

The `.value` migration distinguishes two cases per flax's deprecation notice:

• `self.iterations` and `self.training_steps` wrap `jnp.zeros((), ...)` (Variable[Array]) → use `variable[...]` getter/setter.
• `self._metrics_accumulator` wraps a `PyTreeNode` struct (Variable[non-Array]) → use `.get_value()` / `.set_value()`.

Warning count

• Before: 100 warnings across the suite.
• After: 9 warnings, all upstream and out of scope for this PR:
  • `distrax/_src/utils/math.py` — `jax.core.get_aval` deprecation. Fixed when distrax catches up to a newer JAX.
  • `pgx/core.py` — `player_id` argument to `observe` deprecated. Replacement isn't obvious and the test sites genuinely need per-player observation; deferred for a focused PR if/when pgx provides a migration path.

Test plan

• `uv run pytest tests/` — 233 passed, 2 skipped, 9 warnings.

🤖 Generated with Claude Code