Add tool calling support and expand model catalog

[cpetersen]

Summary

Adds tool/function calling support to the red-candle RubyLLM provider, enabling local models to invoke tools through RubyLLM's standard tool interface.

New: Tool Calling

• supports_functions? now returns true
• New Tools module bridges RubyLLM::Tool ↔ Candle::Tool
• perform_tool_completion! routes tool-bearing requests through red-candle's chat_with_tools with execute: false (RubyLLM manages the execution loop)
• Generated UUIDs for tool call IDs (red-candle doesn't generate them)
• Default max_length: 1000 for tool calls (Qwen3 needs room for <think> blocks)
• Compatible with RubyLLM 1.10 through 1.14 (handles tool_prefs/thinking kwargs)

Verified

chat = RubyLLM.chat(model: "MaziyarPanahi/Qwen3-4B-GGUF")
chat.with_tool(PriceLookup)
response = chat.ask("What is the price of a widget?")
# => Tool called: price_lookup(product: "widget")
# => "The unit price of a widget is $9.99."

Expanded Model Catalog (22 models)

Added all models supported by red-candle 1.5:

• Qwen3 (0.6B, 4B), SmolLM2 (360M), Yi-1.5, Granite (7B, 4.0 Micro), GLM-4
• Phi-3 GGUF, Phi-4 GGUF, Mistral v0.3
• TinyLlama safetensors

Dependencies

• red-candle ~> 1.5 (requires tool calling support)
• ruby_llm >= 1.10 (tested compatible through 1.14)

Test plan

• 129 plugin specs pass, 0 failures
• Provider registration works across RubyLLM 1.10-1.14
• Chat generation works across RubyLLM 1.10-1.14
• Tool calling verified end-to-end with Qwen3-4B through RubyLLM

🤖 Generated with Claude Code