Add challenge 96: INT8 KV-Cache Attention (Medium)

[claude]

Summary

• Adds challenge 96: INT8 KV-Cache Attention (Medium difficulty)
• Decode-phase multi-head attention where the KV cache is stored as int8 with per-token float32 scale factors — matching how production LLM serving systems (TensorRT-LLM, vLLM) halve KV-cache memory bandwidth versus fp32
• Solvers must dequantize K and V on-the-fly (K_float[h,s,d] = K_int8[h,s,d] × k_scale[h,s]) then run scaled dot-product attention
• Teaches: INT8 dequantization fused with attention, mixed-precision arithmetic, warp-level reductions, memory bandwidth trade-offs in LLM inference

Test plan

• pre-commit run --all-files passes (black, isort, flake8, clang-format, mojo format)
• run_challenge.py --action run — example test passes
• run_challenge.py --action submit — all functional + performance tests pass on NVIDIA Tesla T4
• Checklist from CLAUDE.md verified: all 6 starter files present, correct comment style for medium difficulty, HTML has all 4 required sections, performance bullet matches generate_performance_test()

🤖 Generated with Claude Code