What does this PR do?

nnx.cached_partial crashes on raw jax.Array / numpy.ndarray attributes. The mechanism:
StaticCache.variables stores flatten() leaves, which can be raw arrays, but two downstream
sites assume every entry is a Variable — the flatten with_paths=False fast path calls
get_raw_value() on each cached entry (the reported crash), and the unflatten cache-update
loop calls update_from_state/set_raw_value (a latent crash on the output path).

This implements the read-only behavior discussed in the issue: raw array leaves are passed
through as-is when flattening, and the merge update loop skips non-Variable entries, so
Variable propagation is unchanged. The cached_partial docstring now documents the read-only
semantics, and the StaticCache.variables annotation is widened to match.

No warning is emitted when a cached function would have updated a raw array: comparing
returned leaves to cached arrays by identity always differs after jit, and by value would
force device syncs, so the constraint is documented instead.

Adds a regression test with raw jax.Array and np.ndarray attributes alongside an
nnx.Variable counter that must keep receiving updates across cached calls;
tests/nnx/transforms_test.py and tests/nnx/graph_utils_test.py pass on CPU.

Fixes #5109

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). nnx.cached_partial errors on jax Arrays not wrapped in nnx.Variable #5109
• The documentation and docstrings adhere to the documentation guidelines.
• This change includes necessary high-coverage tests. (No quality testing = no merge!)