fix: keep entity verticle requests local when a local instance exists

[ikalachy]

Problem

Entity verticles register a consumer at the shared FQN address EntityVerticle[<FQN>] which forwards to the co-located DataVerticle[<name>] instance. In a clustered setup where the same entity verticle runs on multiple nodes, both addresses are registered on every node, so a request could be round-robined to a remote instance.

When the remote instance replies with an EntityWrapper for an entity type whose OData model is not buffered on the requesting node, EntityWrapperMessageCodec.encodeToWire fails, the reply is never delivered, and the caller times out after 30s — surfacing as an intermittent HTTP 500.

This was observed in production on the calendar service: TimelineEvents intermittently 500'd because its nested ProductMaping entity request round-robined cross-node and the (unmodeled) EntityWrapper reply could not be serialized back.

Fix

Two changes in AbstractEntityVerticle:

1. Inner forward is local-only — the FQN consumer forwards to ownAddress with setLocalOnly(true). Its only job is to hand off to the co-located instance; it must never round-robin across the cluster.
2. FQN address registered as a local consumer — on deployment the EntityVerticle[<FQN>] address is added to NeonBee's local-consumer registry (and removed on stop()), so local-preferred requests resolve to the local instance.

Test

EntityVerticleLocalPreferredClusterTest uses the in-JVM FakeClusterManager (no external infra, CI-safe). Deploys a node-tagging entity verticle on both nodes and asserts requests with a local instance always resolve locally. Verified it fails without the fix (30s reply timeout on DataVerticle[...]) and passes with it.