feat: Model Arena — side-by-side model comparison

[rexlunae]

Inspiration: Unsloth

Unsloth's Model Arena lets you compare any two models side-by-side with the same prompt. This is especially powerful after fine-tuning to compare base vs. fine-tuned outputs, or when choosing between models for a task.

What RustyClaw could do

Add a model comparison feature accessible via the TUI/desktop client:

1. Side-by-side chat: Send the same prompt to two different models (or two different providers) simultaneously, display responses next to each other
2. Diff mode: Highlight differences between responses
3. Blind test mode: Randomize which is A/B, reveal only after user votes
4. Model comparison tool: An agent tool (model_arena) that lets agents themselves compare model outputs
5. Latency/token tracking: Show side-by-side stats for each model

Scope

• TUI client: split-pane view using ratatui
• Desktop client: side-by-side panels in Dioxus
• Gateway: support multi-model concurrent queries (already partially supported via session spawning)
• Core: a compare_models tool for agents

Why this matters for RustyClaw

RustyClaw is provider-agnostic and encourages model experimentation. A built-in comparison tool makes that experimentation concrete rather than requiring users to manually copy-paste between sessions.

Prior art

• Unsloth: https://unsloth.ai/docs/new/studio/chat.md#model-arena
• Chatbot Arena (LMSys) popularized the blind comparison format
• Many chat UIs have this (OpenRouter, Poe, etc.)