Fix memory leak in bounded ReplaySubject due to retaining the node index

[akarnokd]

indefinitely once the Subscriber caught up.

For issue #1865.

[benjchristensen]

This test failed both times:

rx.internal.operators.OperatorMergeDelayErrorTest > testErrorInParentObservableDelayed FAILED
    org.mockito.exceptions.verification.TooLittleActualInvocations: 
    stringObserver.onNext("hello");
    Wanted 2 times:
    -> at rx.internal.operators.OperatorMergeDelayErrorTest.testErrorInParentObservableDelayed(OperatorMergeDelayErrorTest.java:516)
    But was 1 time:
    -> at rx.observers.TestObserver.onNext(TestObserver.java:78)
        at rx.internal.operators.OperatorMergeDelayErrorTest.testErrorInParentObservableDelayed(OperatorMergeDelayErrorTest.java:516)

No idea why or if it's related, but I'd like to understand before merging this.

[akarnokd]

I don't think MergeDelayError uses ReplaySubject.

I don't fully understand the OperatorMerge, but it contains a lot of mutable variables and synchronized blocks and I'm not sure but I found some anomalies in:

https://github.com/ReactiveX/RxJava/blob/1.x/src/main/java/rx/internal/operators/OperatorMerge.java#L505

It synchronizes on the MergeProducer's this but reads the MergeSubscriber's wip field which generally is accessed through the MergeSubscriber's this. Same seems to be true for the scalarValueQueue.

[benjchristensen]

I've tested things locally and it's working for me.

[benjchristensen]

I opened an issue to track the issue you brought up regarding merge: #1885