Add challenge 91: MoE Token Dispatch (Medium)

[claude]

Summary

• Adds challenge 91: MoE Token Dispatch (Medium difficulty)
• Teaches stable parallel scatter — given T tokens and their expert assignments, pack each token's feature vector into the correct per-expert buffer [E, capacity, D], preserving the original token ordering within each expert's group
• Natural sequel to challenge 67 (MoE Top-K Gating): that challenge picks which expert each token goes to; this challenge dispatches the tokens there
• Requires solvers to think about: parallel histogram (atomic counts), work distribution across heterogeneous token loads, and stable scatter without race conditions

Challenge design

• Input: x [T, D] float32, expert_idx [T] int32 (0 … E-1), scalar dims T, D, E, capacity = T
• Output: dispatched_x [E, capacity, D] float32, token_counts [E] int32
• Ordering constraint: within each expert's buffer, tokens appear in ascending original-index order (stable dispatch)
• Difficulty: Medium — empty solve stub for all 6 frameworks
• Performance test: T = 16,384, D = 512, E = 8 (fits well within 16 GB T4 VRAM)

Test plan

• All 6 starter files present (.cu, .pytorch.py, .triton.py, .jax.py, .cute.py, .mojo)
• 10 functional test cases: single token, empty experts, one-per-expert, skewed distribution, power-of-2 sizes, non-power-of-2 sizes, realistic size, zero inputs
• Validated with run_challenge.py --action submit → ✓ All tests passed
• pre-commit run --all-files passes (black, isort, flake8, clang-format, mojo format)
• Checklist in CLAUDE.md fully reviewed

🤖 Generated with Claude Code