Fix Rx serialization bug in takeUntil again and the concurrent issue in BufferUntilSubscriber

[zsxwing]

Sorry that #1656 is not a correct fix.

[zsxwing]

Also fixed BufferUntilSubscriber as per #1677

[akarnokd]

I thought about this but couldn't find any reason the original had race if the caller to onXXX is serialized proprly, as you mentioned in #1677.

[zsxwing]

In the original codes, https://github.com/zsxwing/RxJava/blob/af4fc88b3321247c6e6e2778ae6f3226c742dac0/rxjava-core/src/main/java/rx/internal/operators/BufferUntilSubscriber.java#L101, L101-L105 is not atomic. It means between L101(checking if the buffer is empty) and L105(setting PassThruObserver), other threads may add new items into the buffer. Assume that after L105(setting PassThruObserver), no further message will come. Therefore, drainIfNeededAndSwitchToActual won't be called, and the items in the buffer won't be taken.

[zsxwing]

I thought about this but couldn't find any reason the original had race if the caller to onXXX is serialized proprly, as you mentioned in #1677.

Sorry that my comments in #1677 about PublishSubject confused you. It's irrelevant to this PR.

[akarnokd]

Thanks, that makes sense.

[benjchristensen]

Yeesh that was painful. Thanks both of you.