Halfstream receive window is created afresh after reactor refactor

!126 (merged) broke the previous behaviour of the HalfStream created when a stream is terminated retaining the receive window from when the stream was alive (since the windows are now singlethreaded instead of shared, and shipping the window over to the reactor on closure isn't completely trivial). This means that an attacker could send slightly more data than they should be able to after a stream closes, but it is still bounded (they can't send more than RECV_WINDOW_INIT cells).

We should either fix this, or conclusively determine that it isn't a security risk.