Avoid unnecessary memory allocation for covariance downdate in SGPR prediction strategy (#2559)

* Added ability for priors of transformed distributions to have their p… (#2551) (#2552)

* avoid unnecessary memory allocation for downdate in SGPR prediction

---------

Co-authored-by: David Eriksson <[email protected]>

Author: JonathanWenger

gpytorch/models/exact_prediction_strategies.py

@@ -865,5 +865,5 @@ def exact_predictive_covar(self, test_test_covar, test_train_covar):
                 "This is likely a bug in GPyTorch."
             )
 
-        res = test_test_covar - (L @ (covar_cache @ L.transpose(-1, -2)))
+        res = test_test_covar - MatmulLinearOperator(L, covar_cache @ L.mT)
         return res