feat(qwen3.6): add Qwen3.6-35B-A3B-NVFP4 + MTP Atlas recipe

[tbraun96]

Summary

Adds the Atlas recipe for RedHatAI/Qwen3.6-35B-A3B-NVFP4 with the upstream-bundled MTP K=2 draft head. Mirrors the existing qwen3.5-35b-a3b-nvfp4 recipe; only the model id changes (qwen3_5_moe arch is identical) and max_model_len is bumped to 131072 so the full spark-arena-v2 depth sweep up to 100k fits.

Result (live-validated on a single GB10)

sparkrun benchmark run recipes/qwen3.6/qwen3.6-35b-a3b-nvfp4-atlas.yaml --profile spark-arena-v2 --port 8888

• Decode @ tg=128, pp=2048, depth=0, concurrency=1: 214.59 tok/s (4.66 ms TPOT)
• vs current Spark Arena leaderboard #6 (Qwen3.6-35B-A3B-NVFP4 on vLLM): 77.07 tok/s → 2.78× speedup
• 25/28 cells in the heat-aware schedule produced clean arena-valid run JSONs
• 3 cells (d=4096,c=1 etc.) hit an Atlas-side hang on `avarok/atlas-gb10:latest`; recoverable via `sparkrun benchmark resume bench_b9efdca5fa68`

Reproducibility

• Image: `avarok/atlas-gb10:latest` (stock public)
• Model: `RedHatAI/Qwen3.6-35B-A3B-NVFP4` (stock public)
• Sparkrun: v0.2.31 (PyPI)
• No engine patches required

Test plan

• `sparkrun show recipes/qwen3.6/qwen3.6-35b-a3b-nvfp4-atlas.yaml` parses
• `sparkrun benchmark run … --profile spark-arena-v2` produces consolidated.json with 25 cells
• Result beats every NVFP4 entry on the current Spark Arena leaderboard

🤖 Generated with Claude Code