[classification] Migrate inference pipeline from Ray to Fray v2

Replace all Ray-specific code in classification/inference with Fray v2
equivalents. File-level parallelism now uses fray.v2.current_client().submit()
with JobRequest instead of @ray.remote + ray.wait/ray.get. The autoscaling
actor pool creates Fray actors via client.create_actor() and polls futures
with ActorFuture.result(timeout=0) instead of ray.wait. AutoClassifierRayActor
is removed; ping() and classify() are now methods on AutoClassifier directly.
ray.util.queue.Queue replaced with stdlib queue.Queue.

Fixes #4088

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: github-actions[bot]