feat: Enhance tool error handling to allow agent recovery

[hiromesh]

feat: Enhance tool error handling to allow agent recovery

Summary

This PR enhances the tool error handling mechanism to allow agents to detect and respond to tool call failures instead of having executions terminate on errors.

Changes

• Add is_error flag to ToolCallOutputItem class to indicate when a tool call failed
• Modify run_single_tool function to catch exceptions and return error information instead of propagating exceptions
• Ensure error information is included in agent's conversation history for LLM context
• Add comprehensive test coverage for error handling functionality

Benefits

• Agents can now continue execution after tool failures
• LLMs can learn from error messages and attempt alternative approaches
• Improves agent resilience by allowing recovery from transient tool errors
• Enables more sophisticated error handling patterns in agent workflows

Testing

Added a new test case that verifies:

• Tool errors are properly caught and flagged
• Error messages are correctly returned to the agent

This change is backward compatible and all existing tests pass.