Speed up NNX graph flattening by avoiding unnecessary array-ref checks

[KookiesNKareem]

This PR reduces Python overhead in NNX graph flattening by avoiding unnecessary is_array_ref checks in common non-ref paths.

Specifically:

• _graph_flatten skips variablelib.is_array_ref(node) when the node is already known to be a graph node, pytree node, or Variable.
• variablelib.is_array_ref fast-paths runtime Ref objects and concrete non-ref arrays, while preserving the existing jax.typeof fallback for AbstractRef and traced array cases.

This keeps mutable-ref behavior unchanged while reducing repeated calls into JAX abstract-value logic on the hot path.

Perf on a warmed no-op nnx.jit graph-overhead workload under jax==0.10.0:

• Baseline median: 1089.38 us/call
• Patched median: 936.28 us/call
• Improvement: ~14.1%

Checklist

• This PR fixes a minor issue (e.g.: typo or small bug) or improves the docs (you can dismiss the other checks if that's the case).
• This change is discussed in a Github issue/discussion (please add a link).
• The documentation and docstrings adhere to the documentation guidelines.
• This change includes necessary high-coverage tests. (No quality testing = no merge!)

Validation:

• tests/nnx under jax==0.10.0: 1686 passed, 10 skipped
• tests/nnx under jax==0.10.2: 1686 passed, 10 skipped
• Focused mutable-ref/variable/graph utility suite against the applied checkout under jax==0.10.0: 239 passed, 9 skipped
• Verified is_array_ref(AbstractRef) preserves previous behavior.