Eval denormalization bug: using (z + MIN) * (MAX - MIN) instead of z * (MAX - MIN) + MIN

[William-island]

In evaluation, inverse min–max scaling is implemented as:

real = (event_time_non_mask[:,0,:].detach().cpu() + MIN[1]) * (MAX[1]-MIN[1])
gen = (sampled_seq[:,0,:1].detach().cpu() + MIN[1]) * (MAX[1]-MIN[1])

and

real = (real + torch.tensor([MIN[2:]])) * (torch.tensor([MAX[2:]])-torch.tensor([MIN[2:]]))
gen = (gen + torch.tensor([MIN[2:]])) * (torch.tensor([MAX[2:]])-torch.tensor([MIN[2:]]))

It shifts and scales wrongly, biasing MAE/RMSE and distance metrics.

For min–max normalization z = (x - MIN) / (MAX - MIN), denormalization is x = z * (MAX - MIN) + MIN