fix: address CodeQL and coderabbit review comments

Remove unused variables (actual_nall, threshold) flagged by CodeQL.
Use trainer._optimizer_step() instead of trainer.optimizer.step() in
the dynamic-shapes test to match the real training path.

Author: Han Wang

source/tests/pt_expt/test_training.py

@@ -204,7 +204,7 @@ def test_compiled_handles_varying_nall(self) -> None:
                     lr = trainer.scheduler.get_last_lr()[0]
                     _, loss, _more_loss = trainer.wrapper(**inp, cur_lr=lr, label=lab)
                     loss.backward()
-                    trainer.optimizer.step()
+                    trainer._optimizer_step()
 
                     # Loss should be a finite scalar
                     self.assertFalse(torch.isnan(loss))