[XLA:GPU] Fix implementation of Thunk::ShouldBlockFutureThunks.

The old implementation did not do what it was intended to do.  Instead
of making the next thunk on each other stream wait for this thunk to
complete, it made all following thunks on this stream wait for this
thunk to complete.  (Which is redundant, since streams are serialized
anyway.)

PiperOrigin-RevId: 179938424