Dead lock protection

[idelvall]

Several deadlock scenarios can occur:

1. Submitted job with excessive RSS promise (and schedulerCfg.maxJobRSSBytes=-1)
2. Submitted child job (job launched by other job that waits for it) and the sum of the promises from the root ancestor to it exceeding schedulerCfg.maxTotalRSSBytes
3. All running jobs waiting for queued jobs to end

The two first are irresolvable scenarios, so submission should fail.

The last one can be addressed in different ways.

• Making the scheduler preemptive, so running jobs could be stopped for later execution. A way of freezing a process is needed for this. This would allow to freeze a process to make room to other process descendants (not its own descendants)
• Canceling running jobs, maybe introducing the concept of job idempotence for later relaunching