Add challenge 73: Bilateral Filter (Medium)

[claude]

Summary

• Adds challenge 73: Bilateral Filter (Medium difficulty)
• The bilateral filter is an edge-preserving image smoother: for each pixel it computes a weighted average of its neighbors, combining a spatial Gaussian (proximity) with a range Gaussian (intensity similarity), so sharp edges are preserved while flat regions are smoothed
• This is fundamentally different from the existing Gaussian blur (challenge 28), which uses a fixed precomputed kernel — bilateral filter weights are data-dependent and change per pixel

What this teaches

• Data-dependent stencil weights: unlike convolution, the filter kernel varies at every pixel based on local intensity differences
• 2D stencil with halo memory regions: efficient solutions tile the image into shared memory blocks with border halos
• Per-pixel normalization: each output pixel divides by the sum of its weights (non-linear, cannot be expressed as a simple convolution)
• Algorithm design tradeoffs: naive O(N²R²) per-pixel loops vs. shared-memory tiling to amortize global memory bandwidth

Test plan

• 10 functional tests covering: single pixel, 2×2 zeros, 3×3 edge case, 4×4 negatives, 16×16 and 64×64 power-of-2, 100×100 and 255×255 non-power-of-2, 512×512 and 1000×1000 realistic sizes
• Performance test: 2,048×2,048 image with radius=5
• Validated CUDA solution passes on NVIDIA TESLA T4 via run_challenge.py --action run
• All 6 framework starters present (CUDA, PyTorch, Triton, JAX, CuTe, Mojo)
• Pre-commit linting passes (black, isort, flake8, clang-format)
• Checklist in CLAUDE.md verified

🤖 Generated with Claude Code