Add best-effort tool-calling/template auto-detection

[michalharakal]

Summary

Use metadata to infer likely chat template and possible tool-calling support, without treating it as authoritative.

Scope

• Inspect GGUF metadata such as tokenizer.chat_template when available
• Inspect tokenizer/config-side chat template hints where available
• Add best-effort inference logic for tool-calling compatibility
• Support explicit override when metadata is absent or wrong

Acceptance criteria

• Auto-detection can suggest a likely template/provider
• Explicit configuration still wins over auto-detection
• Missing metadata does not break existing behavior