fix: Fix a leak in PrefixAndTail operator.

[He-Pin]

Motivation:
Fix @queimadus's report in #1566

Modification:
When the upstream finishes quickly but the downstream is not pulled, a leak can happen, so clean up that.

Result:
I think the leak should be fixed now

code from @queimadus

import org.apache.pekko.actor.ActorSystem
import org.apache.pekko.stream.scaladsl.{Sink, Source}
import org.apache.pekko.util.ByteString

import scala.concurrent.Await

object PekkoQuickstart extends App {
  private implicit val system: ActorSystem = ActorSystem()

  val s = Source
    .repeat(())
    .map(_ => ByteString('a' * 400000))
    .take(1000000)
    .prefixAndTail(50000)
    .flatMapConcat { case (prefix, tail) => Source(prefix).concatLazy(tail) }

  val r = Source.empty
    .concatAllLazy(List.tabulate(30000)(_ => s): _*)
    .runWith(Sink.ignore)

  Await.result(r, scala.concurrent.duration.Duration.Inf)
  println(r.value)

//  Source
//    .repeat(s)
//    .take(30000)
//    .flatMapConcat(x => x)
//    .runWith(Sink.ignore)
//    .onComplete(println(_))

//  Source.empty
//    .concatAllLazy(List.tabulate(30000)(_ => Source.lazySource(() => s)): _*)
//    .runWith(Sink.ignore).onComplete(println(_))
}

before:

after:

Seems there are more leak points.

I think the problem is the current logic design where arrays are been used to track and then with the current deign, there will be many many instances of the logic.

So That problem is harder to fix than we expected, but this pr is still valid.

[He-Pin]

@queimadus continue here #1652

[pjfanning]

lgtm - but would prefer if this was not yet backfit to the 1.1.x branch

We've had problems with changes in the stream internals causing unforeseen issues.

[He-Pin]

Okay, let me remove the backport label from it