fix: Add tool_calls array to OpenAI chat completions response

[advenimus]

Summary

This PR fixes a bug where the convert__anthropic_message_to_openai_chat__response function was not converting Anthropic tool_use blocks to OpenAI tool_calls format in chat completion responses.

The Problem:
When Claude returns a tool use response, the converter correctly set finish_reason: "tool_calls" but the actual tool_calls array was missing from the response message:

{
  "choices": [{
    "finish_reason": "tool_calls",
    "message": {
      "role": "assistant",
      "content": ""
      // MISSING: "tool_calls": [...]
    }
  }]
}

The Fix:
This PR adds handling for tool_use blocks in the response converter, converting them to the OpenAI tool_calls format:

{
  "choices": [{
    "finish_reason": "tool_calls", 
    "message": {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "toolu_xxx",
        "type": "function",
        "function": {
          "name": "get_gmail",
          "arguments": "{}"
        }
      }]
    }
  }]
}

Use Case

This fix is required for n8n AI Agent workflows and other OpenAI-compatible clients that rely on the tool_calls array to execute function calls. Without this fix, the AI Agent sees the tools are available but Claude's tool call requests are lost in translation.

Changes

• Added handling for tool_use blocks in convert__anthropic_message_to_openai_chat__response
• Converts Anthropic tool_use format (id, name, input) to OpenAI tool_calls format (id, type, function.name, function.arguments)
• Sets content to None when there's no text content (per OpenAI spec)
• Added json import for serializing tool arguments

Test plan

• Tested with CCProxy-API v0.2.0 on Docker
• Verified tool_calls array is now present in responses
• Tested with n8n AI Agent workflow - tools now execute correctly

[CaddyGlow]

Reused a shared helper to build OpenAI tool_calls for Anthropic adapters (streaming + non-streaming) to ensure consistent JSON serialization of function.arguments. Added unit coverage for tool_use -> tool_calls and strengthened the sample-based test on claude_messages_tools to assert tool names and JSON args.

Tests:

uv run pytest tests/unit/llms/formatters/test_anthropic_to_openai_helpers.py; 
uv run pytest tests/unit/llms/formatters/test_formatter_endpoint_samples.py -k anthropic_message_to_openai_chat.