Notes from convo with @mattjj – this is an OK fix to the vmap issue, but doesn't really get to the root of the problem (e.g. the same thing occurs with jvp of an identity function.

That said, the result of transform(lambda x: x)(np_array) is a corner case that may not be all that important in practice. However, the issue of transform(jnp.asarray)(np_array) returning a non-JAX array is more surprising. There are a couple ways we could address this:

1. convert numpy arrays on input (as in this PR)
2. convert numpy arrays on output
3. ensure that jnp.asarray lowers to some primitive – probably something similar to the existing copy_p.

@mattjj things (1) or (2) are fine, with a slight preferance toward (2). (3) may be better, but it's hard to predict whether it may have downsides we're not anticipating. I'm going to explore (3) a bit before moving forward here.