Add challenge 106: Token Embedding Layer (Medium)

[claude]

Summary

• Adds challenge 106: Token Embedding Layer — the BERT-style input embedding kernel that gathers token and positional embeddings, sums them, and applies LayerNorm with learnable affine parameters.
• Exercises indirect-addressed gathers from two large lookup tables, coalesced reads across the embedding dimension, and per-token reductions in a single fused kernel — solidly Medium, distinct from the existing attention/sampling/quantization challenges.

Validation

• Verified end-to-end on the live platform via scripts/run_challenge.py (--action submit → All tests passed on Tesla T4).
• pre-commit run --all-files clean.

Test plan

• Linting passes
• Reference impl matches the CUDA solution across all 9 functional cases and the perf case (B=32, T=512, V=30,000, P=2,048, D=768)
• First HTML example matches generate_example_test()
• Performance test memory ~150 MB (fits 5× in 16 GB Tesla T4 VRAM)

🤖 Generated with Claude Code