docs(pd_disaggregation): add AMD MORI and Mooncake recipes

[Lzy17]

Summary

Adds AMD/ROCm recipes to the PD Disaggregation page, mirroring the existing NVIDIA + Ascend layout.

• Mooncake (AMD): Llama-3.1-8B single-node 1P1D + DeepSeek-R1 multi-node 1P1D (TP8 per role), with the ROCm + AITER env preamble and bnxt_re* / ionic_* (AMD AI-NIC) variants.
• MORI: new section with install notes, the same two shapes, and the cross-node KV transfer requirements (--privileged, /dev/infiniband, all local bnxt_re* in --disaggregation-ib-device).
• Includes the SGLANG_USE_AITER_AR=0 workaround for the post-2026-04-10 ROCm aiter all-reduce regression, and the bnxt_re kernel-ABI mitigation note.

Locally verified: Llama 1P1D single-node on both transports; cross-node MORI handshake + warmup KV transfer over bnxt_re0 between two MI355 nodes.

Test plan

• mint dev preview renders new sections in TOC
• Single-node MORI 1P1D launches and serves
• Single-node Mooncake 1P1D launches and serves
• Cross-node MORI handshake + warmup transfer (full 2GB MR pin is node-specific bnxt_re driver behavior, not recipe-related)