Unexhaustive list of structural improvements

[adematti]

desilike development started when jax wasn't so widespread. But desilike now heavily relies on JAX. I think we should bite the bullet and structure desilike in a more jax-friendly way:

• all data structures, including Calculator-derived classes, should be pytrees, following Equinox ideas. Then everything will go through standard jax jit's and vmap's nicely, and every JAX-based sampler.

• support for multi-dimensional parameters, and allow for Samples to be split row-wise (in case the parameter space is really large...).

• make the Sampler and Profiler API a bit nicer, if we just want to swap in a different sampler, and keep the export to Samples the same.
  e.g. a function make_sampler(run=likelihood -> array of samples for varied parameters) would return a Sampler

• for non-JAX functions, one could use callbacks

I'm not keen on relying on numpyro at the moment (our Parameter class works well), but we should certainly implement a to_numpyro() method.