Fix min-snr-gamma for v-prediction and ZSNR.

[feffy380]

Based on the discussion in #673
Matches the reference implementation from the paper's author.

Note: With this PR, you should use --min_snr_gamma without --scale_v_pred_loss_like_noise_pred

The old implementation with apply_snr_weight and scale_v_prediction_loss_like_noise_prediction splits the calculation into two steps:

loss *= min(snr, gamma) / snr
loss *= snr / (snr+1)

The problem here is that with --zero_terminal_snr it is possible for the first step to divide by zero.
This PR simplifies the weight calculation to min(snr, gamma)/(snr+1) to avoid dividing by zero.
This makes scale_v_prediction_loss_like_noise_prediction unnecessary, but I did not remove it yet to avoid breaking existing workflows and GUIs. Maybe a deprecation notice should be added to it?

[kohya-ss]

Thank you for this! I didn't fully catch the discussion in the issue, but I believe this works fine. I will merge this sooner. Sorry for waiting a long time.