docs: correct batch endpoint, feature store, and traffic splitting claims

[apurvapatkeshwar]

Summary

Follow-up in the same spirit as #1788 (Ray Tune / autoscaling corrections): three more places where the getting-started docs describe capabilities that are on the roadmap but not in the code yet.

1. overview.md - the ML workflow mapping table says Batch Inference maps to "Deployment to batch endpoint". There is no batch endpoint resource anywhere in proto/ or go/ (grep for it is empty), and the roadmap doesn't list one either. Offline inference today is a hand-written Uniflow task using Ray - that's what python/examples/llm_prediction/ does. The row now says Uniflow tasks with Ray for offline inference.

2. faq.md - "Yes, Michelangelo AI integrates with feature stores" contradicts the roadmap, which lists the entire Feature Store section (online + offline store, drift monitors, lineage) under Planned. No feature-store integration exists in code; the only trace is a forward-looking feature_store_features schema field in model.proto that references a Feature resource type that doesn't exist. The answer now says not yet, it's on the roadmap, and describes what you can do today (data prep pipelines, data catalog).

3. glossary.md - the Deployment entry claims "traffic splitting and canary rollout" and the Endpoint entry claims "traffic splitting across model revisions (e.g., for A/B testing)". In code, strategies/strategy.go has exactly one strategy case (rolling), and TrafficRoutingActor adds a path-based route per deployment (/cluster/{inferenceServer}/{deployment}) - there is no weighted split or canary anywhere. The roadmap itself lists "Endpoint traffic splitting and shadow routing" and "Deployment rollout strategies (Blast, Zonal, Shadow/A-B)" as Planned. Both entries now describe rolling rollout / name-based routing and defer splitting to the roadmap.

Why it matters

Same rationale as #1788: users evaluating a migration (from SageMaker/Vertex especially, where batch transform and canary endpoints are table stakes) will make decisions off these three sentences. The roadmap already tells the honest story; this aligns the other docs with it.

Testing

• 3 files, +4/-4 lines, prose only - no code changes.
• Verified each claim against code at current main before rewording: no batch endpoint type in proto/go, no feature-store implementation, single rolling strategy case, path-based (not weighted) traffic routing.
• Relative link style (./roadmap.md) matches the existing links in faq.md.