feat: add SimpleStorage payload transfer abstraction

[Ginray]

Summary

• add the narrow PayloadTransfer boundary for SimpleStorage large payloads
• keep ZMQ as the default payload path and control plane
• provide an optional UCX Tagged Host-memory implementation for RoCE RDMA
• preserve Mooncake, YuanRong, RayStore, and existing backend ownership
• discover node-local RDMA/GID/netdev capabilities without machine-specific NIC names
• keep the MR focused on implementation and necessary automated tests

Why

SimpleStorage needs a small data-plane boundary for large encoded payloads without changing its
ZMQ control protocol or forcing other storage backends onto the same transport. UCX is opt-in and
does not replace the storage backend or control plane.

User impact

The default remains:

payload_transfer: zmq

Users can select ucx when TQ is built with UCX support and each participating Ray node has a
matching RoCE-v2 path. UCX uses RC for the payload data lane; TCP may be used only as the UCX
wireup/auxiliary lane when the environment requires it.

Validation

• ruff check . passed
• git diff --check passed
• protocol and lifecycle tests are retained in tests/test_payload_transfer.py
• official, unmodified UCX v1.22.0 was validated separately on A2-26/A2-27
• SimpleStorage cross-node UCX PUT/GET/CLEAR passed in the recorded validation

The current native close path can leave a server-side Connection reset by remote peer warning
after a client force-closes; the data and CLEAR checks still pass.