Use filelock to build extension_device backend one at a time (#137930)

Fixes https://github.com/pytorch/pytorch/issues/136125
Fixes https://github.com/pytorch/pytorch/issues/137026
Fixes https://github.com/pytorch/pytorch/issues/137027

The compilation fails during `setUpClass`, so disabling the test doesn't do nothing.  The theory I have for this flaky issue is that `test_open_device_registration` from both `TritonExtensionBackendTests` and `ExtensionBackendTests` are run in parallel and cleaned up while the other is still in fly, causing flaky failure.

Here is an example failure https://github.com/pytorch/pytorch/actions/runs/11331105492/job/31512603585#step:22:1710

Pull Request resolved: https://github.com/pytorch/pytorch/pull/137930
Approved by: https://github.com/malfet

test/inductor/test_extension_backend.py

@@ -22,6 +22,8 @@ except ImportError:
         ExtensionWrapperCodegen,
     )
 
+from filelock import FileLock, Timeout
+
 import torch._inductor.config as config
 from torch._inductor import cpu_vec_isa, metrics
 from torch._inductor.codegen import cpp_utils
@@ -51,10 +53,20 @@ TestCase = test_torchinductor.TestCase
 class BaseExtensionBackendTests(TestCase):
     module = None
 
+    # Use a lock file so that only one test can build this extension at a time
+    lock_file = "extension_device.lock"
+    lock = FileLock(lock_file)
+
     @classmethod
     def setUpClass(cls):
         super().setUpClass()
 
+        try:
+            cls.lock.acquire(timeout=600)
+        except Timeout:
+            # This shouldn't happen, still attempt to build the extension anyway
+            pass
+
         # Build Extension
         torch.testing._internal.common_utils.remove_cpp_extensions_build_root()
         source_file_path = os.path.dirname(os.path.abspath(__file__))
@@ -77,6 +89,10 @@ class BaseExtensionBackendTests(TestCase):
 
         torch.testing._internal.common_utils.remove_cpp_extensions_build_root()
 
+        if os.path.exists(cls.lock_file):
+            os.remove(cls.lock_file)
+        cls.lock.release()
+
     def setUp(self):
         torch._dynamo.reset()
         super().setUp()