Add challenge 70: Merge Sorted Arrays (Medium)

[claude]

Summary

• Adds challenge [LeetGPU Challenge Problem] Add 1D RMS Norm #70: Merge Sorted Arrays (Medium difficulty)
• Given two sorted float32 arrays A (M elements) and B (N elements), produce a single sorted output array C of M+N elements
• The sequential merge has a serial dependency chain; an efficient GPU solution must use the merge path / co-ranking algorithm — each thread independently binary-searches to find its diagonal on the merge path and merges a local tile without any inter-thread dependencies

Why this is interesting

This challenge exercises GPU concepts that are underrepresented in the current set:

• Independent work decomposition at the thread/block level for an inherently sequential algorithm
• Binary search on GPU to compute co-ranks
• Merge path algorithm (Odeh et al., 2012) — a classic GPGPU primitive
• Memory access patterns for irregular work distribution

It is not element-wise at all; every output element's value depends on a non-trivial combination of positions in both input arrays.

Test plan

• 13 functional test cases covering: edge cases (M=1/N=1), all-zero inputs, all-negative inputs, interleaved ranges, non-overlapping ranges, power-of-2 sizes (16, 64/128, 256/512), non-power-of-2 sizes (30/25, 100/73, 255/200), realistic sizes (1K/500, 5K/3K)
• Performance test: M = N = 10,000,000 (80 MB output, well within 16 GB VRAM × 5)
• All 6 framework starters present (CUDA, PyTorch, Triton, JAX, CuTe, Mojo)
• black, isort, flake8 --bugbear, clang-format all pass
• Challenge loads via importlib (module name + signature verified)

🤖 Generated with Claude Code

[kunal-mansukhani]

@claude Rebase and bump this pr to be challenge_id 71