feat: implement cheap clone via Arc<Config> for all LLM backends

[aniketfuryrocks]

Summary

This PR wraps backend configurations in Arc to enable O(1) cloning instead of expensive deep copies. All backend structs now derive Clone.

• Wrap config fields in Arc<*Config> for all backends
• Make config structs and fields public for direct access
• Make client field public on all backends
• Add with_client() constructors for HTTP client injection/pooling
• Add doc comments to all config struct fields

Backends Updated

Standalone backends:

• Anthropic, Google, Ollama, DeepSeek, XAI, AzureOpenAI, Phind, ElevenLabs

OpenAI-compatible backends (via OpenAICompatibleProvider):

• OpenAI, Groq, Mistral, Cohere, OpenRouter, HuggingFace

Usage

// Clone is now cheap (Arc increment)
let provider2 = provider.clone();

// Direct field access
let api_key = &provider.config.api_key;
let client = &provider.client;

// HTTP client injection for connection pooling
let shared_client = reqwest::Client::new();
let provider = Anthropic::with_client(shared_client, "api-key", ...);

Testing

• All 63 unit tests pass
• All 12 doc tests pass
• Clippy passes

---

Draft assisted with OpenCode

[aniketfuryrocks]

@graniet quickly drafted this PR with opencode, let me know how it looks. I tried to make minimal changes. I think we need to refactor the backend directory to correctly couple related code in the future. This is a follow up pr with my comments on #100

@graniet @nazq

[graniet]

Thank you @aniketfuryrocks for your work, I will review soon !

[graniet]

LGTM !