Add challenge 90: Causal Depthwise Conv1d (Medium)

[claude]

Summary

• Adds challenge 90: Causal Depthwise 1D Convolution (Medium difficulty)
• The solver must implement a batched depthwise causal 1D conv over a (B, L, D) tensor, where each channel d is convolved with its own kernel weight[d, :] and only past/present positions are accessible (zero-padded boundary)
• This is the short convolution used in Mamba/SSM models before the selective scan, making it a real-world inference building block
• Teaches: channels-last memory layout, sliding-window boundary conditions, no cross-channel mixing (depthwise structure), and the tradeoff between tiling along L vs. D

Key design decisions

• Convention: weight[d, 0] applies to the current position, weight[d, K-1] to the oldest — natural convolution notation (reference impl flips before calling F.conv1d)
• Performance test: B=8, L=2048, D=4096, K=4 (Mamba-like dimensions, ~512 MB total, fits 5× in 16 GB)
• Difficulty: Medium — empty solve stub only, no kernel scaffold

Test plan

• Validated locally: example test passes with reference CUDA solution on NVIDIA TESLA T4
• All 6 starter files present and correct
• pre-commit run --all-files passes (black, isort, flake8, clang-format, trim whitespace)
• Checklist in CLAUDE.md verified: HTML starts with <p>, uses <h2> sections, example matches generate_example_test(), performance bullet matches generate_performance_test()

🤖 Generated with Claude Code