[regression] Artifact download hangs upon transfer failure

There's still a hole in the parallel artifact resolution we introduced in MNG-3379 for a scenario like this:

• request to resolve multiple artifacts from the same groupId (i.e. all of these artifacts will be resolved by the same thread)
• any but the last artifact of the group fails to resolve

Upon error, the existing code will not try to resolve other artifacts in the same group, thereby never decrementing the countdown latch which guards the execution of the main thread.