[nnx] add support for standalone Variables

[cgarciae]

What does this PR do?

Adds support for using Variables outside of graph nodes, meaning they can now be passed directly to transforms, including inside pytrees without a parent graph node, and the ability to call split, merge, and update on them directly.

rngs = nnx.Rngs(0)
w = nnx.Param(jax.random.normal(rngs(), (2, 3)))
b = nnx.Param(jnp.zeros((3,)))
count = nnx.Variable(jnp.array(0))

@nnx.jit
def linear(w, b, count, x):
  count += 1
  return x @ w + b[None]

x = jax.random.normal(rngs(), (1, 2))
y = linear(w, b, count, x)

assert count.value == 1
assert y.shape == (1, 3)