Fix a purity breakage in unconsAsync, resulting in wild behaviour

[SystemFw]

Anything involving unconsAsync (e.g. merge, interruptWith) was affected, and the side-effecting behaviour resulted in weird things happening. As an example:

object Bug {
  import fs2._
  import cats._, implicits._
  import cats.effect._

  import scala.concurrent.ExecutionContext.Implicits.global
  import scala.concurrent.duration._

  def stepAsync(s: Stream[IO, Unit]): Stream[IO, Unit] =
    s.pull.unconsAsync.flatMap { ap =>
      ap.pull.flatMap {
        case None => Pull.done
        case Some((hd, tl)) => Pull.output(hd) >> tl.pull.echo
      }
    }.stream

  def stream = Scheduler[IO](2).flatMap { s =>
    val a = s.sleep_[IO](1.seconds) ++ Stream.eval(IO(println("tick")))

    //side effects
    val broken = stepAsync(a)
    def works = stepAsync(a)

    broken ++ broken ++ broken ++ works ++ works ++ works ++ broken ++ broken ++ broken
  }.run.unsafeRunSync
}

all the broken except the first would complete immediately.

I would have liked to add a test for this, but struggling for ideas to make it nice

[avakhrenev]

You could inline this Pull:

Pull.fromFreeC { Algebra.eval[F, Nothing, Promise[F, Res]](async.Promise.empty).flatMap { p =>

[SystemFw]

yeah that's a good idea, I'll do that

[SystemFw]

In fact, the whole operation seems quite similar to start, so maybe it's worth thinking for a while how to make it a bit nicer

[avakhrenev]

About tests: there only 2 tests for Scheduler. Maybe you could add a couple of tests for other Scheduler methods, which also demonstrate this issue?

[SystemFw]

Note that e.g. retry has its own set of tests, and now that I look at it, there's no reason SchedulerSpec should be in jvm and not in shared, and be perhaps unified with RetrySpec (more items for #1009 )

It's probably a good idea to have more tests for that, but to test the behaviour exhibited here I'd have to involve both a scheduler and merge, plus something like repeat... it feels a bit weird testing for referential transparency :)

[pchlupacek]

lgtm. I don't think so we need tests for that a such it was more like escaping the ref transparency bounds.

[SystemFw]

let me reformat real quick