There is no load balancing in your scheme, just failover/redundancy.
Load balancing is needed when you have for example 2 GL nodes and specific Input is running on each of them.
If you have balancer then it can route requests to both nodes.
Without balancer all requests will go to the same node(where you have your floating IP), so one node can be overloaded and second node will do nothing.
So this assumption:
if only the master Graylog node is processing the incoming messages
is not really correct.
Message is processed by the node where Input is running. Master node is handling things related to Indexes, metadata and so on