dsh-llm-sampling: DeepSeek Harness Model Sampling Strategy Plugin

`dsh-llm-sampling` is a DeepSeek Harness plugin designed to enforce sampling policies via the `agent/request` waterfall. It manages deployment-level parameters and supports fields such as `temperature`, `topP`, and `topK`, using configured Profiles to override earlier proposals in the request for fine-grained control over the generation distribution. Installation requires pinning the commit and ensuring pnpm version compatibility. It only injects sampling values and does not affect the KV cache, but it depends on adapter support for the specific fields. The plugin must be activated after named routing; if no route is configured, it remains unchanged.