[FEATURE] Allow to inject a message to the agent mid-work

[edenreich]

Summary

This is useful if the agent went idle too long or is working to long in the wrong direction, as an orchestrator you sometimes wish to change the direction an agent is working.

This feature includes touching also code in the ADK to enable message injecting during execution.

An example use-case, say you asked the browser agent to go ahead and collect some data, then on the final moment after the task is already submitted you wanted to narrow down the results to only 5 datasets - this should be possible to inject it mid-execution.

Acceptance Criteria

• In the "/tasks" overview I'm able to click a task and inject a user message
• The ADK supports message queue that are injected into the context during a task execution after each loop the message can be injected
• Follow-up tickets created if it's in the ADK repository to support it
• It's tested
• It's documented