Raise FullyBayesianPosterior deprecation warning only when used (#2004)

Summary:
Pull Request resolved: #2004

X-link: pytorch/botorch#2116

Currently, this warns on import, which leads to excessive warnings since it is imported in the `__init__` file. Moving to `FullyBayesianPosterior.__init__` resolves the issue.

Reviewed By: Balandat, esantorella

Differential Revision: D51445186

fbshipit-source-id: 58a2137ad7d1b79d391e606e601e600bf35b9656

ax/models/torch/utils.py

@@ -612,7 +612,7 @@ def pick_best_out_of_sample_point_acqf_class(
 def predict_from_model(model: Model, X: Tensor) -> Tuple[Tensor, Tensor]:
     r"""Predicts outcomes given a model and input tensor.
 
-    For a `FullyBayesianPosterior` we currently use a Gaussian approximation where we
+    For a `GaussianMixturePosterior` we currently use a Gaussian approximation where we
     compute the mean and variance of the Gaussian mixture. This should ideally be
     changed to compute quantiles instead when Ax supports non-Gaussian distributions.