feat: Connections hub — unified provider management with model mixing

[rexlunae]

Inspiration: Unsloth

Unsloth's Connections feature provides a polished UX for managing all model providers in one place:

• Cloud providers (OpenAI, Anthropic, OpenRouter) with API keys
• Model servers (llama.cpp, vLLM, Ollama) with base URLs
• Unified model dropdown: Shows local, cloud, and external models in a single sorted list
• Prompt caching indicators: Shows which models support prompt caching
• Per-provider capabilities: Dynamically exposes reasoning levels, image support, thinking modes
• Provider-native features: OpenAI web search, Anthropic code execution passed through

Current RustyClaw state

RustyClaw has a provider registry in providers.rs with support for Anthropic, OpenAI, Google, xAI, OpenRouter, Ollama, and custom providers. But:

• No UI for managing providers beyond the onboard wizard
• No runtime add/remove of providers
• No capability discovery (just hardcoded per-provider)
• Ollama integration exists but is separate from the main provider UX

What RustyClaw could do

1. Connections panel in TUI/Desktop: Add/remove/edit providers with test-connection button
2. Runtime provider CRUD: Add new providers without restarting the gateway
3. Capability detection: Auto-detect model capabilities (vision, thinking, tool use, prompt caching) via provider API
4. Mixed model chat: Let a session switch models mid-conversation (useful for: "this model's bad at math, switch to o1 for this query")
5. Provider health: Show latency, error rate, rate limit status per provider

Scope

• rustyclaw-tui + rustyclaw-desktop: Provider management panels
• rustyclaw-core: Runtime provider add/remove + capability detection
• rustyclaw-core: Provider health tracking (latency histograms, error counters)

Prior art

• Unsloth Connections: https://unsloth.ai/docs/integrations/connections.md
• OpenRouter's unified model interface
• LiteLLM's provider abstraction