Deadlock when process killed

[phargogh]

When a subprocess managed by taskgraph is killed by the operating system, the pool will automatically spawn a new process, but not the events that have been allocated to that process. Because of this, the graph deadlocks waiting on events that will never be triggered.

To reproduce:

1. Create a new graph in multiprocessed mode (n_workers >= 1)
2. Execute a task
3. Kill that task before it completes
4. Observe graph hanging

A practical way to trigger this is to use a memory-constrained environment such as Sherlock. On Sherlock, just make sure we have at least 1 task that uses more memory than we have requested for the SLURM job.

Although I suppose it might be ideal to have the appropriate events recreated so the graph can continue to execute, I think it might be better to simply detect that the process has been terminated and then terminate the graph.