Loosen a tolerance to stop test from failing with CUDA (#2512)

Summary:
Pull Request resolved: #2512

Context:

`botorch.test.mdoels.test_gpytorch.TestModelListGPyTorchModel.test_model_list_gpytorch_model` has been failing when run with CUDA for a couple days, with a numerical error slightly exceeding the set tolerance in single precision.

This PR:
* Slightly loosens the tolerance in single precision
* Updates a couple assertions to use `assertAllClose`
* Updates a couple assertions to use `assertWarns`

Reviewed By: Balandat

Differential Revision: D62343447

fbshipit-source-id: 17f054083f0dfec17b4e036e86c7d80c5ea434c1

test/models/test_gpytorch.py

@@ -437,13 +437,12 @@ def test_model_list_gpytorch_model(self):
             m2 = SimpleGPyTorchModel(train_X2, train_Y2)
             model = SimpleModelListGPyTorchModel(m1, m2)
             self.assertEqual(model.num_outputs, 2)
-            with warnings.catch_warnings(record=True) as ws:
+            msg = (
+                "Component models of SimpleModelListGPyTorchModel have "
+                "different batch shapes"
+            )
+            with self.assertWarnsRegex(Warning, msg):
                 self.assertEqual(model.batch_shape, torch.Size([2]))
-                msg = (
-                    "Component models of SimpleModelListGPyTorchModel have "
-                    "different batch shapes"
-                )
-                self.assertTrue(any(msg in str(w.message) for w in ws))
             # test different batch shapes (not broadcastable)
             m2 = SimpleGPyTorchModel(
                 train_X2.expand(3, *train_X2.shape), train_Y2.expand(3, *train_Y2.shape)
@@ -491,17 +490,15 @@ def test_model_list_gpytorch_model(self):
                 ),
                 dim=1,
             )
-            self.assertTrue(torch.allclose(expected_mean, posterior2.mean))
+            self.assertAllClose(expected_mean, posterior2.mean)
             expected_covariance = torch.block_diag(
                 posterior.covariance_matrix[:2, :2],
                 mt_posterior.covariance_matrix[:2, :2],
                 mt_posterior.covariance_matrix[-2:, -2:],
                 posterior.covariance_matrix[-2:, -2:],
             )
-            self.assertTrue(
-                torch.allclose(
-                    expected_covariance, posterior2.covariance_matrix, atol=1e-5
-                )
+            self.assertAllClose(
+                expected_covariance, posterior2.covariance_matrix, atol=1e-5
             )
             # test output indices
             posterior = model.posterior(test_X)
@@ -521,7 +518,8 @@ def test_model_list_gpytorch_model(self):
                 self.assertAllClose(
                     posterior_subset.variance,
                     posterior.variance[..., output_indices],
-                    atol=1e-6,
+                    atol=2e-6 if dtype is torch.float else 1e-6,
+                    rtol=3e-4 if dtype is torch.float else 1e-5,
                 )
             # test observation noise
             model = SimpleModelListGPyTorchModel(m1, m2)