Refactor tool results to typed output models

[gekinci]

This PR standardizes how code interpreter and generic tool results are represented across streaming, tool execution, and OpenAI message serialization. The motivation is to replace loosely shaped dict/string payloads with typed result models, making tool output handling more predictable, preserving backwards compatibility for legacy streamed content, and avoiding accidental leakage of internal metadata into LLM-facing payloads.
It closes #74.

Changes:

• Added shared Pydantic models for CodeInterpreterResult, CreatedFile, and GenericToolResult.
• Updated code interpreter execution, file detection, and server responses to return typed result objects.
• Updated web search responses to use GenericToolResult.
• Changed streaming variants so CodeOutput and ToolOutput carry typed result models instead of raw dicts/strings.
• Extended normalization helpers for legacy code output and generic tool output payloads.
• Centralized LLM payload serialization for code results, excluding preview_url and url_sent_to_model.
• Preserved image URL forwarding behavior for production while ensuring each created image URL is only sent to the model once.
• Updated tool result parsing to validate typed outputs and set preview URLs on created files.
• Expanded unit tests for typed streaming outputs, legacy normalization, created file handling, OpenAI message conversion, and generic tool result serialization.