In-engine speculative decoding

Configure draft models for faster local generation.

arka speculative vllm --draft-model qwen3:1.7b
arka speculative vllm --draft-model qwen3:1.7b --num-draft-tokens 6 --apply

Arka stores draft settings without claiming that every runtime supports them. The active vLLM, Exo, Ollama, or MLX build must expose speculative-decoding options when launched. Verify generated output quality and latency after enabling it.