Make jnp.asarray lower to asarray_p in the simplest cases.

[jakevdp]

This is an experiment toward a fix to #25745 and #18020.

Basically, we currently lower jnp.asarray to nothing in many cases, and this leads to strange behavior under some transformations; for example:

In [1]: import jax

In [2]: import numpy as np

In [3]: out = jax.vmap(jax.numpy.asarray)(np.arange(4))

In [4]: type(out)
Out[4]: numpy.ndarray

The fix here ensures that asarray lowers to a primitive, so that these kinds of corner cases behave as expected.

Many of the test failures here should be fixed by work related to #25931, which removes calls to jnp.asarray from JAX API implementations.