`fix_noise_scheduler_betas_for_zero_terminal_snr` should come before `prepare_scheduler_for_custom_training`

I think the correct implementation should be like this.
```python
noise_scheduler = DDPMScheduler(
beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000, clip_sample=False
)
if args.zero_terminal_snr:
custom_train_functions.fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler)
prepare_scheduler_for_custom_training(noise_scheduler, accelerator.device)
```
The current implementation calculates `noise_scheduler.all_snr` before even fixing the betas in the scheduler. This would make `apply_snr_weight` use the wrong SNR when training zero terminal SNR models with min SNR gamma.

Locate the training entry point that constructs DDPMScheduler and calls fix_noise_scheduler_betas_for_zero_terminal_snr and prepare_scheduler_for_custom_training. Trace when all_snr is calculated, then verify the zero-terminal-SNR path prepares the scheduler in the intended order so apply_snr_weight uses updated SNR values.

Written by the indexing model from the issue text.