fix: correct operator precedence in lvlb_weights calculation

[Mr-Neutr0n]

Bug

In the lvlb_weights calculation, the expression 1 - alphas_cumprod_prev / 1. - alphas_cumprod is parsed as 1 - (alphas_cumprod_prev / 1.) - alphas_cumprod due to Python operator precedence, instead of the intended (1 - alphas_cumprod_prev) / (1 - alphas_cumprod).

Fix

Added parentheses to ensure correct precedence matching the DDPM paper's posterior variance formula.

[Mr-Neutr0n]

This single-line fix corrects the operator precedence in the x0 parameterization branch of lvlb_weights. Without the parentheses around 1. - torch.Tensor(alphas_cumprod), the subtraction is evaluated after the multiplication, producing (2. * 1) - alphas_cumprod instead of the intended 2. * (1. - alphas_cumprod). The fix is clearly correct and has been open for about three months with no activity. Would a maintainer be able to take a look?