api: move ATTR_ADDRESS_NAME to EquivalentAddressGroup

[shivaspeaks]

Correction to the implementation of gRFC A81-xds-authority-rewriting. The proposal says:

Note that the resolver attribute used here should be a general-purpose one, not something specific to EDS;

The endpoint hostname attribute from gRFC A81 is currently defined in XdsInternalAttributes, which makes it unreachable from other modules.

Move the key to EquivalentAddressGroup alongside the other endpoint attributes and re-export it via InternalEquivalentAddressGroup. XdsInternalAttributes held nothing else, so it is removed and its callers now reference the key directly.

This is needed by the autosharding LB policy gRFC A119, which keys its endpoint map on the A81 hostname. autosharding cannot depend on xds because xds will depend on autosharding.

CC: @sauravzg

[ejona86]

Why does this exist? Can't all callers of this use EquivalentAddressGroup directly?

The point of this class is if there's something we don't want to make public API. So it would be useful if the attribute wasn't in EAG at all.

[shivaspeaks]

Updated in the follow-up: #13076

[ejona86]

Oh, nm. I failed to notice it wasn't public in EAG.

[shivaspeaks]

But then you also said

So it would be useful if the attribute wasn't in EAG at all.

So we can keep it in internal only. What problem will it have?

[shivaspeaks]

What purpose does it serve defining in EAG at all?

[ejona86]

When building with Bazel, we use a two-stage build and build Internal* files separately. So if this attribute was used in io.grpc itself, it can't use the Internal* copy. In this case it isn't used in io.grpc at all so it really doesn't matter, but the normal pattern is to just make Internal* accessors so it could be used in io.grpc.