GAN Losses: Minimax and Non-Saturating

For discriminator logits on real data d_real and on generated data d_fake:

• discriminator loss: $-\mathbb E[\log\sigma(d_{real})] - \mathbb E[\log(1 - \sigma(d_{fake}))]$
• minimax generator loss: $\mathbb E[\log(1 - \sigma(d_{fake}))]$
• non-saturating generator loss: $-\mathbb E[\log\sigma(d_{fake})]$

Implement d_loss(d_real, d_fake), g_loss_minimax(d_fake) and g_loss_nonsat(d_fake) stably from logits (use $\log\sigma(z) = -\text{softplus}(-z)$), plus g_grads(d_fake) returning the gradient of each generator loss with respect to the logits as a tuple (minimax_grad, nonsat_grad) (mean over the batch).