Use --job-dir and not --output-dir

In Cloud ML models, use the supplied --job-dir as the output dir. This avoids the need to do code like this:

output_dir = os.path.join(
output_dir,
json.loads(
os.environ.get('TF_CONFIG', '{}')
).get('task', {}).get('trial', '')
)

(it's already done for --job-dir)