What this PR does / why we need it:

When using RaySource with BatchFeatureView in a KubeRay-managed cluster on Kubernetes, materialization fails with multiple issues:

• auth_token / auth_server are required in kuberay_conf even for in-cluster pods that authenticate via service accounts, making deployment on managed Kubernetes platforms impossible without manual token injection.
• runtime_env cannot be passed to the Ray client connection, so additional pip dependencies (e.g. datasets, sentence-transformers) needed by Ray workers cannot be specified via configuration — they must be baked into every worker image.
• HuggingFace Dataset serialization fails in from_huggingface() with AttributeError: 'Dataset' object has no attribute 'to_arrow'. HF Dataset objects are backed by memory-mapped .arrow files which cannot be deserialized on remote Ray workers across the network.
• datasets package missing from feast[ray] extras, requiring users to manually install it despite RaySource depending on it.