Apply extensive input to extensive output forwarding in scan.

[dfm]

This PR applies a small optimization to scan for the case where some outputs are directly forwarded inputs. For example, in

def body(c, x):
  return c + 1, x

_, y = jax.lax.scan(body, 0, jnp.arange(10))

y would be computed by copying each element of x into a new array. After this change, y is returned as x without making a copy.

[mattjj]

Beautiful, thanks @dfm