CSI volumes can fail for `poststop` tasks for a stopped job after client restart

In https://github.com/hashicorp/nomad/pull/16722 we've fixed a crash described in detail in https://github.com/hashicorp/nomad/issues/16623#issuecomment-1492415089, but the crash fix still leaves poststop tasks in a failed state under the following circumstances:

• a job has a poststop task that wants a CSI volume
• the client stops
• the server stops the job
• the client starts again
• the client restores the allocrunner but the client receives the stop from GetClientAllocs before the allocrunner hits the Prerun alloc hooks.
• this will result in shouldRun returning false, which prevents the CSI hook from running.
• but the poststop task is still run, which means its volume hook doesn't have access to any info from the CSI hook, so the poststop task fails.

It would be nice if we could serialize client startup such that GetClientAllocs happens before we restore the state, because that would at least enforce a reasonable workflow, but that would also cause the bug to always happen instead of just some of the time. Do we even want to guarantee that poststop tasks to survive across client restarts in this way? Are there other lurking cases of this scenario in other hooks that we just haven't found yet?

Start with client/allocrunner/alloc_runner.go, especially shouldRun and the Prerun alloc hooks, then trace the CSI hook and poststop task handling described in the issue. Reproduce the client-restart sequence if possible and establish the intended behavior for poststop tasks and CSI volumes before defining tests and completion criteria.