test(remotagent) deflake a2a cleanup propagation

[wolo-lab]

Summary

• Fix the panic + flakiness in TestA2ACleanupPropagation (agent/remoteagent/v2/a2a_e2e_test.go)
• Join the detached streaming/cancel goroutines so neither logs on t after the test completes.
• Make cancellation-propagation deterministic by cancelling only once the remote subagent's output has round-tripped to the client (i.e. once the remote agent actually knows the subagent task ID).
  Test-only change; no production code is modified.

Background / root cause

The test exercised two failure modes:

1. Panic — Log in goroutine after test has completed. The CancelTask
   goroutine was never joined (cancelResultChan was created but never read),
   so under load it could still be mid-RPC when the test returned, fail on the
   cancelled t.Context(), and call t.Errorf on a finished test.
2. remote cleanup was not called. This was not merely a too-tight
   timeout. Via goroutine dumps + tracing I confirmed the test cancelled the
   parent task right after the first status event — i.e. before the remote
   agent had received any event from the subagent. In that window the remote
   agent doesn't yet know the subagent's (server-assigned) task ID, so
   cleanupRemoteTask returns early (lastEvent == nil) and cannot propagate
   the cancellation. Under GOMAXPROCS=1 / CPU contention this race was lost
   deterministically, so the subagent task kept running and its cleanup never
   fired.

What changed

• sync.WaitGroup + t.Cleanup(wg.Wait) join both detached goroutines before the test tears down (fixes the panic in all exit paths).
• Replaced the single fixed time.After(5s) shared across all cleanup-channel reads with a per-wait, contention-tolerant helper awaitN.
• The test now waits until the subagent's own artifact reaches the client before issuing the cancel. That is the precondition for propagation to be possible, so the (unchanged) propagation assertions become deterministic. Only the moment of cancellation is corrected — the test still exercises the full path.

Note on scope (best-effort propagation)

adk-go's remote-task cancellation propagation (cleanupRemoteTask) is an
additive, best-effort feature; adk-python's remote A2A agent does not cancel
remote tasks at all. Cancelling in the window before the remote agent has
received its first subagent event genuinely cannot propagate (the remote task
ID is unknown to the caller). This PR fixes the flaky test to respect that
contract; making propagation robust in that early-cancel window would be a
separate, behaviour-changing effort (client-assigned task IDs and/or a2a-go
support) and is intentionally out of scope here.