inference: expose opt-in qwen3.6 model policy

[sheawinkler]

Summary

• exposes explicit no-download Dream Mode model policy in /v1/inference/runtime-policy
• recommends Mudler Qwen3.6 GGUF as the opt-in advanced GGUF target for llama.cpp-compatible runtimes
• keeps Huihui abliterated GGUF metadata private-eval gated behind GO_DREAM_ALLOW_UNCENSORED_MODELS=true

Local verification

• GOCACHE=/Volumes/wd_black/codex_go_cache GOMODCACHE=/Volumes/wd_black/codex_gomod_cache go test -count=1 ./... (services/gateway-go)
• bash -n scripts/inference_runtime_policy.sh scripts/inference_backend_guard.sh scripts/benchmark_inference_backends.sh
• python3 -m py_compile scripts/agent/contextlattice-dream
• git diff --check