feat(mockib): RDMA verbs data path (perftest / GPUDirect RDMA)

[giuliocalzo]

Priority: Medium

Context. Full ibv_devinfo (non--l) is intentionally out of scope today because it requires faking the mlx5 uverbs ioctl surface.

Gap. No verbs data path: ibv_rc_pingpong, perftest (ib_write_bw/ib_read_bw), GPUDirect RDMA, nvidia-peermem.

Proposed scope.

• Evaluate feasibility of a verbs-level shim or a loopback/relayed QP path over the existing TCP fabric to support perftest bandwidth runs.
• Document explicitly what remains unsupported if full uverbs faking stays out of scope.
• E2E (if feasible): cross-node ib_write_bw reports non-zero bandwidth.

Why. Enables RDMA bandwidth validation; extends the fabric relay built for ibping.