feat: add sigterm handling for graceful termination

[linxiulei]

When there are system errors on the hosts (e.g hardware errors), the system is given limited timeout to shutdown so likely the checkpoint saving is unable to complete and also unnecessary if other slices have the full checkpoint to save without system errors.

This change will skip checkpoint saving to reduce the termination latency so that the overall recovery time is improved.

[samos123]

Checkpoint saving can be extremely fast with emergency checkpointing. I'm afraid not saving checkpoints may be worse. Especially since we have a custom method deployed to delete lingering pods.

For example for a 70B model, I'm seeing in memory checkpoint saving time to be 7 seconds.

[linxiulei]

Checkpoint saving can be extremely fast with emergency checkpointing. I'm afraid not saving checkpoints may be worse.

This doesn't stop checkpoint saving in other Nodes without hardware faults. The intent of this PR is to make the faulty nodes to terminate and shutdown as soon as possible while other nodes can do the checkpoint saving so that the overall recovery time is reduced. To illustrate:

node0 (without fault): running -> JobSet restart -> checkpoint saving -> pod terminate -> job creation -> pod start
node1 (with fault): running -> JobSet restart -> checkpoint saving -> pod terminate -> node shutdown for repair -> job creation -> pod start

After this PR:

node 1 (with faulty): running -> JobSet restart -> pod terminate -> node shutdown for repair -> job creation -> pod start

Since node 1 is likely having the longest recovery time for whole workload to be back running, skipping its checkpoint saving step would reduce overall recovery time. Besides that, if there is hardware fault such as link error, the checkpoint saving may fail anyway.