Add challenge 97: Softcap Attention (Medium)

[claude]

Summary

• Adds challenge 97: Softcap Attention (Medium) — multi-head self-attention with tanh logit soft-capping applied to pre-softmax scores
• Models a real inference kernel used in production LLMs such as Gemma 2 (softcap=50) and Grok-1.5; complements existing attention variants (MHA, causal, ALiBi, GQA, sliding-window, linear) without duplicating any merged challenge or pending PR
• Solver must fuse softcap * tanh(scores / softcap) into the scaled dot-product attention pipeline before the row-wise softmax

Test plan

• pre-commit run --all-files passes on all new files
• challenge.py imports cleanly; generate_functional_test() returns 10 cases (edge 1–4, zero input, mixed negatives, power-of-2, non-power-of-2, realistic)
• generate_performance_test fits comfortably within 16 GB VRAM (N=2048, d_model=1024, h=16 → ~32 MB of tensors; kernel uses ~9 KB shared memory per block)
• Validated end-to-end on live platform via scripts/run_challenge.py ... --action submit with a T4 CUDA solution: All tests passed (sample + functional + performance)
• Starter files for all six frameworks compile/run but do not produce correct output

🤖 Generated with Claude Code