UCT/D2P: Multi-channel support with per-EP DPA address and RING_DB flag (#11701)

Replace qp_idx with ep_dpa_addr in descriptor so the DPA poster can
locate the per-channel QP from the EP's DPA heap base plus channel index.
Add UCT_IB_D2P_FLAG_RING_DB so the GPU controls doorbell ringing per-op
based on UCT_DEVICE_FLAG_NODELAY. Split gpu_ep into per-channel and
per-EP structs; pass channel_id through the device API dispatch.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>