Allow method="basinhopping" in find_MAP and fit_laplace

[jessegrabowski]

Closes #454

This PR lets users use the basinhopping algorithm to fit PyMC models. This is in the class of simulated annealing algorithms, except that it also runs an inner optimization loop to find the nearest minimum after proposing a new starting position.

API is pretty clean, if I do say so myself:

    with pm.Model() as m:
        mu = pm.Normal("mu")
        sigma = pm.Exponential("sigma", 1)
        pm.Normal("y_hat", mu=mu, sigma=sigma, observed=rng.normal(loc=3, scale=1.5, size=100))

        optimized_point = find_MAP(
            method="basinhopping",
            use_grad=True,
            use_hess=False,
            use_hessp=True,
            progressbar=True,
            gradient_backend="pytensor",
            compile_kwargs={"mode": "JAX"},
            minimizer_kwargs=dict(method="trust-ncg"),
        )

Gives:

  Process                            Elapsed   Iteration   Step    Target   Objective     ||grad|| 
 ──────────────────────────────────────────────────────────────────────────────────────────────────
  Basinhopping   ━━━━━━━━━━━━━━━━━   0:00:02   100/100     0.617   0.500    186.98814     0.00000
  Minimize       ━━━━━━━━━━━━━━━━━   0:00:00   8/8         0.000   0.000    34964.56397   0.00000

The top progress bar shows the overall algorithm, the current step, the target step size, current best objective function, and the gradient at that current best. The bottom shows progress on the inner optimization loop, so it gets updated a lot. You can see here that the final inner optimization went somewhere stupid, so the objective value is very high.

[jessegrabowski]

I don't think the CI failure is our fault here, seems like a dask bug. I opened an issue on their repo.

[AlexAndorra]

Super neat, thanks @jessegrabowski ! Should I merge despite the CI failure then?

[jessegrabowski]

I'm going to try version pinning dask first