docs: add Inference section to README

[igor-susic1]

Adds an ## Inference section to README.md between the existing Models and Tags sections, documenting where and how kimchi runs model inference.

Covers what's actually implemented in src/ollama.ts and wired in src/cli.ts:

• Default inference against kimchi's hosted LLM infrastructure via the setup API key
• Native local inference through Ollama — silent, non-blocking startup probe; stale ollama provider block removed when the server is unreachable
• Ollama endpoint precedence: → → http://localhost:11434
• OpenAI-compatible surface (${host}/v1) and zero-cost reporting for Ollama models
• Role integration — discovered models appended to builder, reviewer, and explorer pools only; orchestrator/planner/judge/researcher intentionally left untouched
• Tier derivation by parameter size (<8B light, 8–<30B standard, ≥30B heavy, unknown→standard)
• Observability via NODE_DEBUG=kimi:ollama

Every claim is backed by the corresponding source in src/ollama.ts / src/cli.ts; no behavior is changed.

• Documentation update