feat: multi-provider LLM architecture (Provider trait, registry, fallback, /models redesign)

[chinkan]

Summary

• Provider trait + registry: New src/provider.rs with Provider trait, ProviderRegistry, and ProviderConfig — resolves provider/model strings to the correct provider
• Three provider implementations: OpenRouterProvider (with Kimi tool-call fallback + param sanitization), OpenAICompatibleProvider (generic LM Studio/vLLM/llama.cpp), OllamaProvider (delegates chat to OpenAI-compatible, custom /api/tags discovery)
• Config extensibility: New [[provider]] TOML sections, [fallback] chain, backward-compat auto-synthesis from legacy [openrouter] section
• LlmClient refactor: Now a thin routing layer over ProviderRegistry — no direct config dependency
• Agent integration: Fallback chain in process_message loop, per-provider supports_vision flag, set_model persists to correct TOML section
• /models redesign: Multi-step inline keyboard flow — provider selection → model discovery/selection with qualified IDs (e.g., ollama/llama3.1)
• docs: [[provider]] and [fallback] examples in config.example.toml
• Tests: 14 provider registry + 5 config backward-compat tests (393 total pass)

Test Plan

• cargo check — clean
• cargo clippy -- -D warnings — clean
• cargo test — 393 unit + 16 integration, 0 failures