feat: add L1/MAE, Hinge, and Focal losses

[ulises-jeremias]

Only MSE, BCE, CrossEntropy, Huber, KL, NLL, SigmoidCE, and SoftmaxCE exist.
Add three widely used loss functions.

Task

Following the pattern of nn/loss/mse.v and nn/loss/huber.v:

1. Add nn/loss/l1.v — Mean Absolute Error (L1 loss)
2. Add nn/loss/hinge.v — Hinge loss for binary SVM-style training
3. Add nn/loss/focal.v — Focal loss for imbalanced classification
4. Add cases to nn/loss/loss_test.v

Verification

• Loss values match PyTorch for same inputs
• Backward gradients are correct (test with gradcheck if available)