clamp variances to be positive in get_infeasible_cost

Summary:
The posterior variance can be slightly negative due to numerical issues.
This can lead to NaNs in the infeasible cost.
This clamps the variance in the computation to be non-negative to avoid this.

(I guess the alternative would be to clamp this directly in the `posterior` method)

Reviewed By: danielrjiang

Differential Revision: D14846840

fbshipit-source-id: bbfebfbcd60c38afb4b107812ff7ab39b91843e3

Author: Balandat

botorch/acquisition/utils.py

@@ -113,6 +113,6 @@ def get_infeasible_cost(
         The infeasible cost M value.
     """
     posterior = model.posterior(X)
-    lb = objective(posterior.mean - 6 * posterior.variance.sqrt()).min()
+    lb = objective(posterior.mean - 6 * posterior.variance.clamp_min(0).sqrt()).min()
     M = -lb.clamp_max(0.0)
     return M.item()