As per the AWS docs, the managed node group update behaviour would ensure adding new nodes first before deleting the existing one. It also says that by default the node will be drained and respect the PDBs. Have you found it to be different in practice? In any case, I am also using Terraform so will use the lifecycle hack you suggested.

I am going to run Kafka in Kubernetes with Cluster autoscaler enabled so will definitely use Auto Scaling groups for each AZ.