[BUG] Spark jobs should only select UP nodes in current subcluster at the time the job starts

[bryanherger]

This is an issue emerging with v12 Eon mode and sandbox support. Current behavior appears to be to select all nodes; this doesn't work with sandboxing, where some nodes can't write to the main communal storage, and also may break during dynamic scaling, especially when a subcluster shrinks, which means some workers will point to a down node. The correct behavior is to check for Eon mode when the job starts, then select UP nodes in the current subcluster for workers to connect to.