Call os.fsync on flush.

[dchao34]

We encountered this issue while writing tensorboard events into a GCS Fuse mounted directory tree.

It turns out that GCS fuse requires the os buffers to to be flushed before others that also mount the same GCS bucket can see the writes (c.f. https://github.com/googlecloudplatform/gcsfuse/blob/master/docs/semantics.md#readwrites).

Since https://github.com/lanpa/tensorboardX/blob/master/tensorboardX/record_writer.py#L192 only does a python file buffer level flush, one needs to add another line just below that to call os.fsync. Something like this

def flush(self):
        self._writer.flush()
        os.fsync(self._writer.fileno())

With this change, gcs fuse mounted directories behaves as you would expect.

torch.utils.tensorboard has the behavior on GCS fuse described here out of the box.