[FEATURE]: Document that CPU-only local inference is unsuitable for agent workflows

Feature hasn been suggested before.

• I have verified this feature I am about to request has not been suggested before.

Describe the enhancement you want to request

When users configure a local model (Ollama, llama.cpp, etc.) as their opencode provider, there is no guidance about the performance characteristics expected for agent use. Users may spend weeks tuning a CPU-only setup before realizing it is fundamentally too slow.

What I want: Documentation (or a startup check) that warns users when:

• A local provider is configured with CPU-only inference (no GPU acceleration)
• Prefill times exceed a reasonable threshold (e.g. >10s)
• The model has non-disableable thinking/reasoning that adds latency

Concrete suggestion: A note in the docs or a dms doctor-style check:

"CPU-only local models (Ollama without GPU) typically have 30-50s prefill times, making them impractical for agent workflows. Consider using a remote provider or ensuring GPU acceleration (CUDA/ROCm) is configured."

Why it matters: I tested Qwen 30B MoE and Qwen 7B on CPU (Ryzen 7 3800X, 32GB DDR4) — both had 30-50s prefill, and the 30B had non-disableable thinking. After removing Ollama entirely and switching to remote inference, the workflow became usable. This guidance would save others the same trial-and-error.

Environment: CachyOS (Arch), no GPU for local inference (AMD RX 6750 XT not used for LLM), Ollama removed.