Harden job lifecycle tracking

[nefrathenrici]

Job status tracking is currently fragile and relies on symbols being passed around.
Clear fragility issues:

• Using PBS, job_status tries two qstat formats and returns :RUNNING on any error, silently masking failures as running jobs.
• Using Slurm, empty squeue output maps to :COMPLETED, conflating "job finished" with "job ID no longer tracked by the scheduler".
• Failure states are asymmetric: PBS returns :FAILED; Slurm does not.

We should:

• store the job status as an enum instead of a symbol
• store full iteration results in a struct
• harden job_status calls, potentially dispatching on the backends

mutable struct IterationResult
    iter::Int
    member_states::Vector{JobState}   # one entry per submitted ensemble member
    n_failed::Int
    n_rerun::Int
end