FR: Separate compile/run methods for pathfinder

[mike-lawrence]

I'm using Pathfinder to generate inits for a subsequent run of NUTS, and for each of my N NUTS chains, I run a separate multi-path pathfinder:

n_chains = int(os.cpu_count()/2)
init_vals = []
seed = 0
while len(init_vals) < n_chains:
	try:
		seed += 1
		with inference_model:
			pathfinder_trace = pmx.fit(
				method = "pathfinder"
				, random_seed = seed
				, num_paths = 5 # more than 5 causes a crash!
				, num_draws_per_path = 1000
				, importance_sampling = 'psis'
				, num_draws = 1
			)
		pathfinder_trace.posterior = pathfinder_trace.posterior.squeeze("chain", drop=True)
		point = {
			var: pathfinder_trace.posterior[var].isel(draw=0).values
			for var in pathfinder_trace.posterior.data_vars
		}
		init_vals.append(point)
	except Exception as e:
		print(f"Error occurred: {e}")
		print("Retrying with a different seed...")

# now we can use the pathfinder trace to initialize the chains
with inference_model:
	posterior_trace = pm.sample(
		chains = n_chains
		, cores = n_chains
		, initvals = init_vals
	)

However, with the current API, there is an unnecessary re-compilation of the model for every run of pathfinder, so it would be useful to have separate compile() and run() methods.