Add vmap reset() regression test for nnx.metrics (#5483)

[chenkuanliao]

What does this PR do?

This adds a regression test for the nnx.metrics behavior reported in #5483.

nnx.metrics.Average.reset() (and Welford.reset()) re-zero their state by assigning a scalar (jnp.array(0, ...)). When a metric is built under nnx.vmap, its state carries a leading batch axis of shape (N,). On the reporter's versions (flax 0.12.0 / jax 0.7.2) the scalar assignment replaced the whole array, collapsing (N,) to (), so a later vmapped update failed with:

ValueError: vmap was requested to map its argument along axis 0, which implies that its rank should be at least 1, but is only 0 (its shape is ())

While investigating I found the crash no longer reproduces on main: the Variable state layer has since changed so a full-index assignment broadcasts a scalar into the existing array in place instead of replacing it, so reset() preserves the batch axis today. I confirmed it still reproduces on flax 0.12.0 / jax 0.7.2 and is gone on main independent of any single commit.

Since there is no longer a live crash to fix on main, this PR is scoped as a regression test only, to lock in the current vmap reset() behavior so it can't silently break again.

What's in this PR

• tests/nnx/metrics_test.py — test_vmap_reset_preserves_shape: construct a MultiMetric(loss=Average('loss')) under nnx.vmap (state shape (N,)), call reset(), assert the state stays shape (N,), then run a vmapped update and assert compute()['loss'] returns the expected per-batch values.

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). — metrics does not work well with vmap #5483
• The documentation and docstrings adhere to the documentation guidelines.
• This change includes necessary high-coverage tests. (No quality testing = no merge!)