Cleanup persistent cuBLASLt workspaces before test_mempool_ctx_multithread

Cherry-pick of #3372

Co-authored-by: Dmitry Nikolaev <139769634+dnikolaev-amd@users.noreply.github.com>

Author: rocm-repo-management-api-6[bot]

test/test_cuda.py

@@ -6713,6 +6713,7 @@ def test_mempool_expandable(self):
 
     @serialTest()
     def test_mempool_ctx_multithread(self):
+        torch._C._cuda_clearCublasWorkspaces()
         torch.cuda.empty_cache()
         segments = torch.cuda.memory._snapshot()["segments"]
         self.assertEqual(len(segments), 0, "Expected empty pool in the beginning")