Don't remove all non-data/non-master nodes during upgrade/scaling

[david-kow]

Currently, we allow all ingest nodes to be unavailable during upgrade and scaling. Upgrade respects maxUnavailable setting and after #1292 is fixed scaling will as well.

Nevertheless, I'd consider an unexpected behavior to remove all ingest nodes making cluster unable to index even if in line with maxUnavailable setting. We should at least leave a single ingest node (like we treat masters) and maybe have even stricter bound (50%?) as going to a single node may cripple cluster capabilities.