Normalize rewards within each GRPO group using its mean and standard deviation, including zero-variance handling.
