[BUG] A potential data corruption in Pandas UDFs

[firestarman]

Thanks to @GaryShen2008 finding this.

There is a potential data corruption in Pandas UDFs using BatchQueue to combine the input and python output batches.
Currently the BatchProducer and BatchQueue use different locks to protect the batch pulling from the input iterator and the batch appending to the queue separately.
So in a two-threaded Python ruuner, there is a race when the reader thread and the writer thread append batches to the batch queue.

One possible case is:

1. the writer thread gets a batch A, but next it pauses.
2. then the reader thread gets the next Batch B, and appends it to the queue.
3. the writer thread resumes and appends batch A to the queue.

Therefore, batch A and B have the reversed order in the queue now, leading to data
corruption when doing the combination.