[C++]: Ability to feed and fetch tensors while keeping them in device…

… memory

when using Session::RunCallable().

PiperOrigin-RevId: 202234757

tensorflow/core/BUILD

@@ -3911,13 +3911,13 @@ tf_cc_test(
     ],
 )
 
-tf_cc_test(
+tf_cuda_cc_test(
     name = "common_runtime_direct_session_test",
     size = "small",
     srcs = ["common_runtime/direct_session_test.cc"],
+    args = [] + if_cuda(["--heap_check=local"]),  # The GPU tracer leaks memory
     linkstatic = tf_kernel_tests_linkstatic(),
     deps = [
-        ":core",
         ":core_cpu",
         ":core_cpu_internal",
         ":direct_session_internal",
@@ -3930,6 +3930,7 @@ tf_cc_test(
         ":test",
         ":test_main",
         ":testlib",
+        "//third_party/eigen3",
         "//tensorflow/cc:cc_ops",
         "//tensorflow/core/kernels:control_flow_ops",
         "//tensorflow/core/kernels:cwise_op",
@@ -3943,8 +3944,7 @@ tf_cc_test(
         "//tensorflow/core/kernels:queue_ops",
         "//tensorflow/core/kernels:session_ops",
         "//tensorflow/core/kernels:variable_ops",
-        "//third_party/eigen3",
-    ],
+    ] + if_cuda([":cuda"]),
 )
 
 # This is identical to :common_runtime_direct_session_test with the addition of