Add challenge 80: Grouped Query Attention (Medium)

[claude]

Summary

• Adds challenge 80: Grouped Query Attention (GQA), a medium-difficulty inference kernel challenge
• GQA is used in virtually all modern LLMs (LLaMA-3, Mistral, Gemma, Phi-3) to reduce KV-cache memory during inference by sharing K/V heads across groups of Q heads
• Solvers must implement scaled dot-product attention where num_q_heads / num_kv_heads consecutive Q heads attend to the same K/V head, requiring correct understanding of memory layout, head grouping, and softmax normalization

Checklist

• challenge.html starts with <p>, has <h2> sections for Implementation Requirements, Example, Constraints
• First example matches generate_example_test() values
• Uses <pre> for 1D data (consistent)
• Constraints includes performance test bullet: num_q_heads=32, num_kv_heads=8, seq_len=1024, head_dim=128
• SVG visualization included (dark theme, #222 background)
• challenge.py inherits ChallengeBase, all 6 methods present
• reference_impl has assertions on shape, dtype, device
• generate_functional_test returns 10 cases covering edge cases, powers-of-2, non-powers-of-2, MQA, MHA-equivalent, zero inputs, realistic sizes
• Performance test fits 5x in 16GB VRAM (~600MB for perf test)
• All 6 starter files present with correct parameter description comments
• Pre-commit lint passes
• Validated with run_challenge.py --action submit — all tests pass on NVIDIA TESLA T4

🤖 Generated with Claude Code