Discovering interpretable elastoplasticity models via the neural polynomial method enabled symbolic regressions