roundrobin: Delegate subchannel creation to pickfirst

[arjan-bal]

This PR switches round round robin to defer to pick first instead of creating and handling SubConns itself. This is part of Dualstack, and makes this a petiole. This PR makes roundrobin a wrapper around the endpoint sharding policy that implements the roundrobin logic.

RELEASE NOTES:

• balancer/roundrobin: Round Robin uses pick first child policies to manage endpoints instead of creating SubConns.
• balancer/roundrobin: Round Robin uses the Endpoints field instead of the Addresses field in resolver state updates.

[codecov]

Codecov Report

Attention: Patch coverage is 89.28571% with 3 lines in your changes missing coverage. Please review.

Project coverage is 82.29%. Comparing base (9afb49d) to head (84556fc).

Files with missing lines	Patch %	Lines
balancer/roundrobin/roundrobin.go	86.36%	2 Missing and 1 partial ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #7966   +/-   ##
=======================================
  Coverage   82.29%   82.29%           
=======================================
  Files         387      387           
  Lines       39065    39067    +2     
=======================================
+ Hits        32147    32152    +5     
- Misses       5582     5602   +20     
+ Partials     1336     1313   -23     

Files with missing lines	Coverage Δ
balancer/weightedroundrobin/balancer.go	83.42% <100.00%> (-0.80%)	⬇️
internal/testutils/balancer.go	83.85% <100.00%> (+1.31%)	⬆️
balancer/roundrobin/roundrobin.go	87.50% <86.36%> (-12.50%)	⬇️

... and 28 files with indirect coverage changes

[dfawley]

This was in the other PR, too.

Probably we should have endpointsharding produce a parsed PF config if it's going to be used frequently?

[arjan-bal]

That's a nice idea. I've raised a PR with this change: #8007

PTAL, I'll rebase both the PRs once #8007 is merged.

[dfawley]

This needed to be done eventually. But maybe for similar things in the future...it could have been easier for this PR (and may still be easier when dealing with future migration PRs) to make the manual resolver's UpdateState and InitialState methods set Endpoints according to the default behavior if it's unset.

[arjan-bal]

This change was not required since the code to populate endpoints based using addresses is present in resolver_wrapper.go:

grpc-go/resolver_wrapper.go

Lines 125 to 132 in f35fb34

	if s.Endpoints == nil {
	s.Endpoints = make([]resolver.Endpoint, 0, len(s.Addresses))
	for _, a := range s.Addresses {
	ep := resolver.Endpoint{Addresses: []resolver.Address{a}, Attributes: a.BalancerAttributes}
	ep.Addresses[0].BalancerAttributes = nil
	s.Endpoints = append(s.Endpoints, ep)
	}
	}

I've reverted changes to tests that use the manual resolver. In tests that call balancer.UpdateClientConnState directly, without creating a real channel, conversion from addresses to endpoints is required.

[dfawley]

This is still necessary, unfortunately, as only balancer.Balancer functions are transferred when embedding.

[arjan-bal]

Reverted.

[arjan-bal]

Closing for now and prioritizing the changes to make Ringhash a petiole policy. Will also make the changes to have the metrics recorder retrieved from a method on the ClientConn as per the offline discussion. Will re-open when I'm back on this.