fix: prevent division by zero in loss when no targets

[Mr-Neutr0n]

Bug

When a batch contains no positive targets, the loss computation in ComputeLoss_NEW divides by zero (via .mean() on an empty tensor), producing NaN loss values. This can happen in two cases:

1. all_loss is empty when no labels exist, causing zip(*all_loss) to fail
2. The validity mask ij has no True entries, causing cat_loss[0][ij].mean() to return NaN

Fix

Added guards to ensure:

• The loss assignment block only runs when all_loss is non-empty
• .mean() is only called when ij has at least one True value
• Objectness loss is still computed against all-zero targets when no positive targets exist, preserving correct gradient flow