…onse (#1736)"

This reverts commit 24a00e6.

we previously returned an empty graphql response at the end of the
response stream to set the `hasNext` field to false, to indicate that no
more responses will come.

That empty response is causing issues in some clients, so 24a00e6 was a
fix to set the `hasNext` on a deferred response from inside query
planner execution, but it does not account for parallel deferred
response executions, so one response might come with `hasNext` to false
then get another one.

This commit attempts another solution, where we go through an
intermediate task that checks if the response stream is closed (it is a
channel, so it implements `FusedStream`). Unfortuantely, right now it
fals to recognize when the stream is closed

since the stream is marked as terminated from inside poll_next, we need
to call it a second time after getting a message, to check if it is
closed, but we cannot do that with an async method, since we need to
send the current message ASAP. So we call `try_next`, and depending on
its result, we either send the current message, or set hasNext on the
current message and send it if the channel is closed, or send the
current message, get the next one and try again to see if there's
another one

there could be a race condition where we consume and send all messages,
then await for the next one, then the stream is closed and we don't have
any message on which we would set `has_next`. So we detect that case and
add a last message