fix: persist tool_result messages to storage during tool processing

[guoyangzhen]

Problem

Only USER input and final ASSISTANT response were saved to chat storage. Tool_result messages created during tool processing were kept in-memory only, so subsequent requests loaded incomplete conversation history.

This caused Bedrock/Anthropic API errors:

tool_use ids found without tool_result blocks immediately after

Root Cause

The orchestrator's save_message only saves the user input and final agent response. The tool_response (USER role with tool results) created by _process_tool_block is added to the in-memory conversation for the next LLM call but never persisted to storage.

Fix

1. Collect tool_response messages in Agent._pending_tool_responses during tool processing loops
2. After the final response is saved, the orchestrator persists these tool_result messages to storage
3. Applied to both Python and TypeScript, both streaming and non-streaming paths

Files changed: 8 files (+66 lines)

• Agent base class (both languages): add pending tool responses array
• BedrockLLMAgent + AnthropicAgent (both languages): collect in processing loops
• Orchestrator (both languages): save pending tool responses after final message

Fixes #396