re-derive sqrt alpha bar and sqrt one minus alphabar

This is the only place these values are ever referenced outside of training code so this change is very justifiable and more consistent.

modules/sd_samplers_timesteps.py

@@ -36,7 +36,7 @@ class CompVisTimestepsVDenoiser(torch.nn.Module):
         self.inner_model = model
 
     def predict_eps_from_z_and_v(self, x_t, t, v):
-        return self.inner_model.sqrt_alphas_cumprod[t.to(torch.int), None, None, None] * v + self.inner_model.sqrt_one_minus_alphas_cumprod[t.to(torch.int), None, None, None] * x_t
+        return torch.sqrt(self.inner_model.alphas_cumprod)[t.to(torch.int), None, None, None] * v + torch.sqrt(1 - self.inner_model.alphas_cumprod)[t.to(torch.int), None, None, None] * x_t
 
     def forward(self, input, timesteps, **kwargs):
         model_output = self.inner_model.apply_model(input, timesteps, **kwargs)