OpenRoom adds llama.cpp, local model support

This is the kind of plumbing that turns a flashy agent demo into something local-model users will actually adopt.

• –Optional API keys lower the barrier for self-hosted, offline, and lab-only setups.
• –Routing through the existing OpenAI-compatible chat path should make llama.cpp fit alongside other local servers more cleanly.
• –Stripping `<think>` blocks is practical UI hygiene for reasoning models that leak chain-of-thought markup.
• –The Reddit example running Qwen3.5-35B-A3B-Q6_K and 27B suggests OpenRoom is already usable with serious local models, not just toy backends.
• –If merged, this gives OpenRoom a stronger pitch as a hackable desktop-agent shell for the LocalLLaMA crowd.