Switch to NLB instead of ELB

[redbaron]

Main benefit: stable IP addresses within VPC, therefore it is expected to be much less broken TCP connections when AWS roll underneath infrastructure

[mumoshu]

@redbaron Thx!
I'm really wanting to switch to NLB - in other words enable NLB by default - rather than adding it as an optional feature.
#598 (comment) reported by @danielfm today strengthened the idea.
WDYT?

Btw, anyway, we need to keep the support for ELB for the China region because NLB has
not available there yet.

[danielfm]

Maybe this should be added as an optional feature via a type (that defaults to classic, or something like that) parameter to apiEndpoints[].loadBalancer; this would allow users to migrate towards the new NLBs without the need to re-create the whole cluster by:

1. Change cluster.yaml to add a new API endpoint of type network, leaving the other endpoint intact
2. Run kube-aws update to create the new load balancer while leaving the old one as-is
3. After making sure the new load balancer is working, change worker[].apiEndpointName to point to it
4. Run kube-aws update to update all workers to use the new endpoint
5. Change cluster.yaml to remove the old load balancer
6. Run kube-aws update to remove the now unused load balancer