No need to wrap inputs in jnp.array

[cgiovanetti]

This wraps all params.get()'s in jnp.array internally, so the inputs to jitted functions are still wrapped in np.array without the user needing to do so in their input params dictionary. I have verified across a couple scenarios (varying LCDM parameters, switching to bbn_type='LINX') that the cached compiled code is still used when non-wrapped float inputs change value.