Run pytorch_core CUDA tests on GPU using TPX

Summary:
Modify contbuild to disable sanitizers, add option to run "cuda" test using TPX RE

(Note: this ignores all push blocking failures!)

Test Plan: CI

Reviewed By: walterddr, cspanda

Differential Revision: D23854578

fbshipit-source-id: 327d7cc3655c17034a6a7bc78f69967403290623

Author: malfet

test/test_cuda.py

@@ -21,7 +21,7 @@
 from torch.testing._internal.common_methods_invocations import tri_tests_args, tri_large_tests_args, \
     _compare_trilu_indices, _compare_large_trilu_indices
 from torch.testing._internal.common_utils import TestCase, get_gpu_type, freeze_rng_state, run_tests, \
-    NO_MULTIPROCESSING_SPAWN, skipIfRocm, load_tests, \
+    NO_MULTIPROCESSING_SPAWN, skipIfRocm, load_tests, IS_SANDCASTLE, \
     slowTest, skipCUDANonDefaultStreamIf, TEST_WITH_ROCM, TEST_NUMPY
 from torch.testing._internal.autocast_test_lists import AutocastTestLists
 
@@ -1732,6 +1732,7 @@ def test_streaming_backwards_device_transfer(self):
         self.assertTrue(b.grad.sum().item() == 4 * size)
 
     @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
+    @unittest.skipIf(not IS_SANDCASTLE, "Does not work on Sandcastle")
     def test_cuda_init_race(self):
         # See https://github.com/pytorch/pytorch/issues/16559
         import subprocess