Add support for capturing agent tool responses

[benjichat]

What's changed

• Added @published public private(set) var agentToolResponses: [AgentToolResponseEvent] = [] property to Conversation class for tracking agent tool responses
• Updated the .agentToolResponse event handler to capture and store tool response events
• Added cleanup logic in resetFlags() and cleanupPreviousConversation() to clear stored responses on reset/disconnect

Why

This extends the previous PR that added support for parsing agent_tool_response events. While the previous PR fixed the parsing errors and handled the events internally, this update exposes the agent tool responses to SDK users through a published property, allowing them to observe and track when agents execute tools.

Usage

SDK users can now observe agent tool executions:

conversation.$agentToolResponses
    .sink { responses in
        for response in responses {
            print("Agent executed tool: \(response.toolName)")
            if response.isError {
                print("Tool execution failed")
            }
        }
    }
    .store(in: &cancellables)

Testing

The existing test coverage for AgentToolResponseEvent parsing validates the underlying functionality. SDK users can now access these events through the published property for their own tracking and debugging needs.