fix(server): Fix bug in flushing in snapshot

[dranikpg]

The bug was here:

  VLOG(2) << "FlushDefaultBuffer " << bytes.size() << " bytes";
  PushBytesToChannel(current_db_, bytes);
  default_serializer_->Clear();

We might block on the channel when pushing. If the snapshot continues serializing, it will write to the default serializer. And once the channel unblocked, the clear will remove everything - not only the part it has written, but also the new one.

The solution would be to make Clear consume only a fixed part - but instead I decided to revert some changes I made previously as well because of the changes in async streamer.

I initially abstracted flushing in snapshot to pushing just bytes taken from the rdb serializer. But since in the async streamer we're gonna make the streamer base to act as a sink as well, we can just use FlushToSink for everyhing, so I removed Flush and make it just PrepareFlush.

[romange]

Do I understand correctly that now we copy everything twice?

[dranikpg]

Do I understand correctly that now we copy everything twice?

No, it works just like before - internal buffer -> string file -> moved out to channel. We had a embedded string file, but it doesn't matter if its owned and moved out or created from scratch, because its filled with one flush.