Fix lr schedule

Call lr scheduler before training so that the update takes place properly also if a device missed out one or several training rounds. This happends e.g. when changing the server device in Swarm SL or neglecting client updates if using a loss threshold. Also added lr scheduler for client for parallel SL.

edml/config/battery/resnet110_cifar100_cost.yaml

+deduction_per_second: 0.02
+deduction_per_mflop: 0.00000005
+deduction_per_mbyte_received: 0.0002
+deduction_per_mbyte_sent: 0.0002