fix: preserve root task input across clarification turns

[vinitsonawane45]

Summary

Fixes root LlmAgent(mode="task") losing its original task input after asking the user a clarification question.

Root cause

Commit ddbb59e changed task-mode root agents to use the generated node path as their isolation scope. This gave root invocations a fixed scope, such as reporting@1, allowing separate root task invocations in the same session to share the same scope.

During task resumption, the original root user message could not be recovered through the existing function-call-based reconstruction because root tasks have no delegating function call.

Fix

• Use target_node_name@ctx.invocation_id as the isolation scope for root task invocations.
• Recover the original root user message for the invocation when reconstructing task input, using the invocation ID encoded in the scope.
• Preserve existing node-path isolation behavior for delegated task nodes.

Tests

Regression coverage includes:

• Root task input preserved across clarification and resumption.
• Sequential root task invocations receive independent isolation scopes.
• Existing peer-task isolation behavior remains covered.
• A2A client/server integration test covering clarification and resumption, including the expected message-role sequence and preservation of the original task input.

Validation

• uv run pytest -v tests/unittests/workflow/test_task_api_e2e.py — 14 passed, 1 xfailed (previously reported).
• uv run pytest tests/unittests/a2a/integration/test_client_server_v1.py -v -k "test_a2a_task_clarification_and_resumption" — 1 passed, 16 deselected.
• uv run pyink --check tests/unittests/a2a/integration/test_client_server_v1.py — passed.
• git diff --check — passed.

Fixes #7384