Python: [A2A] First-class support for `input-required` user input

[SergeyMenshykh]

Context

.NET PR #5743 actually does three things in code:

1. Tracks the A2A TaskState on the session (A2AAgentSession.TaskState, threaded through UpdateSession).
2. In A2AAgent.CreateA2AMessage, branches on typedSession.TaskState == TaskState.InputRequired:
   • InputRequired: sets a2aMessage.TaskId = typedSession.TaskId so the reply is bound to the existing task as input.
   • Otherwise: sets a2aMessage.ReferenceTaskIds = [typedSession.TaskId] for task refinements.
3. Surfaces the input-request message parts on AgentResponseUpdate for TaskStatusUpdateEvent via a new AgentTaskStatusExtensions.GetUserInputRequests() helper that returns the status message''s AIContent parts only when State == InputRequired.

Current Python state

python/packages/a2a/agent_framework_a2a/_agent.py:

• The session has no equivalent task_state tracking; only task_id / context_id are kept.
• _prepare_message_for_a2a (~L621) never sets task_id, even when the prior task was TASK_STATE_INPUT_REQUIRED, so replies are not bound to the in-flight task on the server.
• _updates_from_task_update_event does emit parts from status.message, but it does not gate on INPUT_REQUIRED or signal to the caller that the surfaced content is an input request rather than ordinary intermediate output.

Proposed change

• Track the current TaskState on the A2A session (mirror A2AAgentSession.TaskState).
• In _prepare_message_for_a2a, branch on the previous task state:
  • if INPUT_REQUIRED → set a2a_message.task_id = previous_task_id;
  • else → set a2a_message.reference_task_ids = [previous_task_id] (see the companion "Link follow-up messages via reference_task_ids" issue).
• Add a helper analogous to GetUserInputRequests that returns the status-message parts only when state == INPUT_REQUIRED, and use it when building AgentResponseUpdates for input-required status events so callers can distinguish input requests from regular intermediate content.