Implement Wasserstein Critic Loss

In a Wasserstein GAN, the critic doesn't output a probability — it outputs an unbounded score, trained to score real data high and generated (fake) data low. The gap between those two averages is what the generator is fighting to close.

Task: write wasserstein_critic_loss(real_scores, fake_scores), rounded to 4 decimal places.

• A well-trained critic scores real data higher than fake data, so this loss is negative when the critic is doing its job.
• There's no sigmoid or log here — WGAN scores are raw and can be any real number.