Fix test_ops_error_message.py and run it on CI. (#9640)

This PR fixes #9622, which extracted error messages checks out of
`test_operations.py` into `test_ops_error_message.py`. There were a few
problems with that PR, namely:

- `unittest.main()` wasn't being run: although calling `python -m
pytest` runs it automatically, running it without the `pytest` module
did nothing
- `onlyOnCPU()` would error if `PJRT_DEVICE` environment variable wasn't
set
- `test_ops_error_message.py` wasn't being run on CI

This PR fixes all the aforementioned PRs.

Author: ysiraichi

test/run_tests.sh

@@ -150,6 +150,7 @@ function run_xla_op_tests1 {
   run_dynamic "$_TEST_DIR/ds/test_dynamic_shape_models.py" "$@" --verbosity=$VERBOSITY
   run_eager_debug "$_TEST_DIR/test_operations.py" "$@" --verbosity=$VERBOSITY
   run_test "$_TEST_DIR/test_operations.py" "$@" --verbosity=$VERBOSITY
+  run_test "$_TEST_DIR/test_ops_error_message.py"
   run_test "$_TEST_DIR/test_xla_graph_execution.py" "$@" --verbosity=$VERBOSITY
   run_pt_xla_debug_level2 "$_TEST_DIR/test_xla_graph_execution.py" "$@" --verbosity=$VERBOSITY
   run_test_without_functionalization "$_TEST_DIR/test_operations.py" "$@" --verbosity=$VERBOSITY

test/test_ops_error_message.py

@@ -6,7 +6,7 @@
 
 
 def onlyOnCPU(fn):
-  accelerator = os.environ.get("PJRT_DEVICE").lower()
+  accelerator = os.environ.get("PJRT_DEVICE", "").lower()
   return unittest.skipIf(accelerator != "cpu", "PJRT_DEVICE=CPU required")(fn)
 
 
@@ -158,7 +158,7 @@ def test_mm_raises_error_on_non_matrix_input(self):
     b = torch.rand(2, 2, device=device)
 
     def test():
-      torch.mm(a, b)
+      return torch.mm(a, b)
 
     self.assertExpectedRaisesInline(
         exc_type=RuntimeError,
@@ -172,10 +172,14 @@ def test_mm_raises_error_on_incompatible_shapes(self):
     b = torch.rand(8, 2, device=device)
 
     def test():
-      torch.mm(a, b)
+      return torch.mm(a, b)
 
     self.assertExpectedRaisesInline(
         exc_type=RuntimeError,
         callable=test,
         expect="""mm(): cannot matrix-multiply tensors f32[2,5] and f32[8,2]. Expected the size of dimension 1 of the first input tensor (5) to be equal the size of dimension 0 of the second input tensor (8)."""
     )
+
+
+if __name__ == "__main__":
+  unittest.main()