Only include arg_constraints in pytree_data_fields if they are not lazy_propertys.

[tillahoffmann]

This PR tries to address #1928 to obtain consistent pytree structures.

On a related note, similar behavior is observed with the get_args method. I'm not sure what the "correct" behavior should be in this case.

>>> import jax
>>> from jax import numpy as jnp
>>> from numpyro import distributions as dists
>>> 
>>> 
>>> def make_dist():
>>>     return dists.MultivariateNormal(precision_matrix=jnp.eye(2))
>>> 
>>> 
>>> init = make_dist()
>>> init.get_args()
{'loc': array([0.]),
 'precision_matrix': Array([[1., 0.],
        [0., 1.]], dtype=float32),
 'scale_tril': Array([[1., 0.],
        [0., 1.]], dtype=float32)}
>>> init.covariance_matrix
>>> init.get_args()
{'loc': array([0.]),
 'covariance_matrix': Array([[1., 0.],
        [0., 1.]], dtype=float32),
 'precision_matrix': Array([[1., 0.],
        [0., 1.]], dtype=float32),
 'scale_tril': Array([[1., 0.],
        [0., 1.]], dtype=float32)}

[fehiepsi]

LGTM! Re get_args: it depends on the definition of the method. It might be worth to add a clear description there. Any method with lazy properties might give inconsistent behaviors, not just with get_args.