Fix: scheduler: make multiple-active="stop_unexpected" actually work

The previous implementation covered the scenario in the regression test and not
much else. It unnecessarily added an expected_node member to the native variant
data, when the resource's allocated_to is sufficient to know the expected node.