Support killing tasks

[kayousterhout]

The commit to add support for network and compute monotasks removes support for killing tasks on the executor. This should be added back by communicating relevant information to the LocalDagScheduler, which can avoid executing any new monotasks for the macrotask (and ideally interrupt any running monotasks for the macrotask).

Once we fix this issue, we should be able to fix some residual issues with the unit tests where, when a given unit test is running, sometimes there are old monotasks from previous unit tests still running, causing some weird failures (this should allow us to, in LocalDagScheduler, fail rather than just logging a warning when a monotask that's about to be run isn't in waitingMonotasks).

Once this is done, re-enable the "job group with interruption" test in JobCancellationSuite.scala. This should also significantly speed up the rest of the tests in JobCancellationSuite, which currently take a long time because they wait for tasks to finish executing rather than interrupting the tasks partway through execution.