RGAN: Fix nonsense code and stick to original paper

[HydrogenC]

Fix nonsense code in the training code:

PyTorch-GAN/implementations/relativistic_gan/relativistic_gan.py

Lines 151 to 157 in 36d3c77

	if opt.rel_avg_gan:
	g_loss = adversarial_loss(fake_pred - real_pred.mean(0, keepdim=True), valid)
	else:
	g_loss = adversarial_loss(fake_pred - real_pred, valid)
	# Loss measures generator's ability to fool the discriminator
	g_loss = adversarial_loss(discriminator(gen_imgs), valid)

As shown above, g_loss is assigned and then immediately overwritten, which makes no sense.

Adjustments to better stick to the original paper:

• Update discriminator before generator instead of the opposite
• Revamp loss calculation code (see below)

According to the paper, for RGAN:

For RaGAN:

Implementation reference:
https://github.com/AlexiaJM/RelativisticGAN/blob/master/code/GAN_losses_iter.py#L636-L639