TrainerClient.wait_for_job_status polls until timeout for an invalid or empty status on local backends #877

What happened?

TrainerClient.wait_for_job_status only rejects an invalid status on the Kubernetes backend. On the local process and container backends it is never checked, so a status that can't match just polls until the timeout:

from kubeflow.trainer import TrainerClient
from kubeflow.trainer.backends.localprocess.types import LocalProcessBackendConfig

client = TrainerClient(backend_config=LocalProcessBackendConfig())
job_name = client.train(...)  # any local job
client.wait_for_job_status(job_name, status={"Completed"})  # Spark spelling, TrainJob uses "Complete"
# waits the full 600 s, then TimeoutError

The Kubernetes backend raises ValueError right away for the same call (kubernetes/backend.py, the status.issubset(job_statuses) check). An empty status=set() waits until the timeout on every backend, Kubernetes included, because set() passes the subset check.

I found this while reading the backends side by side. #873 / #874 fix the same thing for SparkClient.

What did you expect to happen?

A ValueError straight away for an empty status or one that isn't a subset of Created, Running, Complete, Failed, as the docstring says ("ValueError: The input values are incorrect"), on every backend.

Environment

Kubeflow SDK from main, Python 3.12. No cluster needed: it reproduces with the local process backend.