2.x: fix scan(seed, f) to emit accumulated values asap

[akarnokd]

For performance reasons, the scan delayed the emission of accumulated items by one which worked well with sources that pushed new items frequently.

Unfortunately, if the source stopped emitting (without completion), the very last accumulated item was stuck and no way to get it out of scan. I considered this type of use being rare but apparently I was wrong.

The PR changes scan to emit accumulated items, including the initial seed, as soon as there is request for it. The operator now uses a stable-prefetch queue-drain found throughout other operators such as observeOn.

Remark: sadly, request trickery such as capturing the first request call, emitting the seed and then requesting n - 1 doesn't work because if the request() call itself happens from another thread and the source is empty, the onNext and onComplete could run concurrently.

Reported in #5089