llm-adaptive: Add an adaptive (automatic routing) model entry point for DSH

llm-adaptive is an MIT-licensed DeepSeek Harness (DSH) plugin designed to solve the challenge of frequently manually switching models within a session due to varying request complexity. The plugin provides an "adaptive (automatic routing)" option in the DSH model selector, enabling intelligent backend routing through per-request complexity classification. Core features include: classifying each request into one of four levels—`low`, `medium`, `high`, or `critical`—using the `deepseek-v4-flash` classifier combined with session context and recent dialogue; supporting Sticky level protection to prevent unnecessary degradation during task execution; and configuring the routing chain entirely via `pool.json`.