Make fully Bayesian imports lazy (facebook#5148)

Summary:

X-link: meta-pytorch/botorch#3265

`import botorch` eagerly loaded Pyro via two import paths:
1. `botorch.fit` → `from pyro.infer.mcmc import MCMC, NUTS`
2. `botorch.models.__init__` → `fully_bayesian.py` → Pyro imports



This diff makes fully Bayesian imports lazy so that `import botorch` no longer
eagerly loads Pyro. The imports are deferred until `fit_fully_bayesian_model_nuts`
is called or `SaasFullyBayesianSingleTaskGP`/`SaasFullyBayesianMultiTaskGP` are
accessed from `botorch.models`.

Differential Revision: D99687047

Author: sdaulton

ax/generators/torch/tests/test_surrogate.py

@@ -51,8 +51,9 @@
 from ax.utils.testing.torch_stubs import get_torch_test_data
 from ax.utils.testing.utils import generic_equals
 from botorch.exceptions.errors import ModelFittingError
-from botorch.models import ModelListGP, SaasFullyBayesianSingleTaskGP, SingleTaskGP
+from botorch.models import ModelListGP, SingleTaskGP
 from botorch.models.deterministic import GenericDeterministicModel
+from botorch.models.fully_bayesian import SaasFullyBayesianSingleTaskGP
 from botorch.models.fully_bayesian_multitask import SaasFullyBayesianMultiTaskGP
 from botorch.models.gp_regression_mixed import MixedSingleTaskGP
 from botorch.models.model import Model, ModelList  # noqa: F401 -- used in Mocks.

ax/utils/testing/tests/test_mock.py

@@ -45,7 +45,7 @@ def test_botorch_mocks(self) -> None:
 
     def test_fully_bayesian_mocks(self) -> None:
         experiment = get_branin_experiment(with_completed_batch=True)
-        with patch("botorch.fit.MCMC", wraps=MCMC) as mock_mcmc:
+        with patch("pyro.infer.mcmc.MCMC", wraps=MCMC) as mock_mcmc:
             with mock_botorch_optimize_context_manager():
                 Generators.SAASBO(experiment=experiment, data=experiment.lookup_data())
         mock_mcmc.assert_called_once()