Add challenge 98: Beam Search Step (Medium)

[claude]

Summary

• Adds a new medium challenge: one decoding step of beam search for autoregressive LLMs.
• Given current cumulative beam log-probs (B, K) and next-token log-probs (B, K, V), expand all K*V candidate continuations and select the top K per batch, sorted by score, returning new beam scores plus parent beam indices and chosen token IDs.
• Exercises broadcast addition, top-K reduction, and structured index decomposition (flat → (parent_beam, token)) — distinct from the existing 1-D top_k_selection because of the per-batch 2-D index space and the dual integer outputs.
• Validated end-to-end on Tesla T4 with a CUDA solution: example test, all functional tests, and the performance test (B=16, K=8, V=50000) pass.

Test plan

• Reference implementation matches example math by hand
• All 10 functional tests have a top-K boundary gap > 1e-3 (no fp32 ties)
• Performance test fits comfortably within 16 GB on T4 (≈25 MB token_logprobs)
• pre-commit run --all-files passes (black, isort, flake8, clang-format, mojo format)
• CUDA solution submitted via scripts/run_challenge.py --action submit → all tests passed

🤖 Generated with Claude Code