Add challenge 85: LoRA Linear (Medium)

[claude]

Summary

• Adds challenge 85: LoRA Linear (Medium difficulty)
• Implements the LoRA (Low-Rank Adaptation) linear layer forward pass: output = x @ W^T + lora_scale * (x @ A^T) @ B^T
• Real-world inference kernel used in every LoRA-adapted LLM deployment
• Teaches two-pass matrix multiplication with a small intermediate (rank-sized) tensor, fused update, and memory optimization opportunities

Test plan

• challenge.py passes all 10 functional test cases locally (validated via run_challenge.py --action submit)
• Example test matches generate_example_test() values
• Performance test fits within 16GB VRAM (batch=256, d_in=d_out=4096, rank=64 → ~74MB total)
• All 6 starter files present and compile/run without producing correct output
• pre-commit run --all-files passes

🤖 Generated with Claude Code