Add challenge 100: Softmax Attention Backward (Medium)

[AaronToh]

#Summary

• Adds challenge 100: Softmax Attention Backward, the backward pass through scaled dot-product attention.
• Given Q (M×d), K (N×d), V (N×d), and upstream gradient dO (M×d), the solver must compute dQ, dK, and dV by propagating gradients through the softmax Jacobian-vector product and the scaling factor 1/√d. This is a natural follow-up to challenge 6 (Softmax Attention).
• Includes challenge.py (8 functional tests covering edge cases, zeros, negatives, powers-of-2, non-powers-of-2, and a performance test of M=512 × N=256 × d=128), challenge.html with full backward pass derivation, and starter files for all six frameworks.