Soft Dice Loss for Segmentation

Segmentation masks are often mostly background, so pixel-wise cross-entropy is dominated by easy pixels. Soft Dice measures overlap directly:

with predicted probabilities $p$, ground truth $g \in \{0,1\}$ and smoothing $s$. Implement dice_loss(p, g, smooth=1.0) over all elements, and dice_grad(p, g, smooth=1.0) returning $\partial L/\partial p$.