parEvalMap and it's variations hang on canceled effect

fs2 version 3.10.2

this code does not terminate:

Stream.unit // number of elements is irrelevant, as long as there are any; does not reproduce on Stream.empty
  .covary[IO]
  .parEvalMap(2)(_ => IO.canceled)
  .compile.drain

while there is no issue with evalMap:

Stream.unit
  .covary[IO]
  .evalMap(_ => IO.canceled)
  .compile.drain

even worse, adding interruptAfter after parEvalMap does not make the code terminate:

Stream.unit
  .covary[IO]
  .parEvalMap(2)(_ => IO.canceled)
  .interruptAfter(5.seconds)
  .compile.drain

it is not necessary for all effects to be cancelled, one is enough. This also hangs:

    Stream(IO.unit, IO.canceled)
      .covary[IO]
      .parEvalMap(2)(identity)
      .compile.drain

expected behavior:

while it's not obvious what to hope for in relation to concurrently executing effects - whether they should also cancel or not - one would expect overall termination semantics to be the same across evalMap and parEvalMap. It is also difficult to reason how to work around such situation.

Reproduce the hang with the parEvalMap and compile.drain examples in the issue, then inspect the parEvalMap implementation and its cancellation handling. Compare its behavior with evalMap, including interruptAfter, and add coverage for a canceled effect; done means these examples terminate with consistent cancellation semantics.