build: upgrade Mooncake MUSA wheel to 0.3.13

[Elon-yf]

Summary

• Upgrade the MUSA Mooncake Transfer Engine wheel from 0.3.12.post1 to
  0.3.13 in the vLLM-MUSA image path.
• Keep the pinned upstream vLLM MooncakeConnector; add only the v0.28
  compatibility needed for MUSA's K/V-first cache layout.
• Keep the upstream blocks-first cache path unchanged.

Validation

Candidate: vllm-musa=a4a81ee4314d9a1d6c1d8e96a2376e9aee13d430
Upstream vLLM: 2cf0a6915ce544dc493a0990f2ea38d81601128a
Driver: 5.2.0-server
Torch/MUSA: 2.11.0.post1+musa5.2.0
Mooncake: 0.3.13

• Focused image/connector/filter contract tests: 25 passed.
• PyPI wheel SHA256:
  f8d54776bd4b66f3c74b33587e30148694105c5792075d3ffd8904184124b3f7.
• Package preflight passed on CPython 3.10/x86_64 with MUSA runtime,
  SUPPORT_MUSA=True, two visible MUSA devices, and no generic CUDA Mooncake
  distribution installed.

Package/runtime evidence:

{"architecture": "x86_64", "device_count": 2, "generic_mooncake_installed": false, "mooncake": "0.3.13", "python": "3.10.12", "support_musa": true, "torch": "2.11.0.post1+musa5.2.0", "torch_musa": "2.11.0.post1+musa5.2.0", "torchada": "0.1.83", "transfer_engine": "<class 'mooncake.engine.TransferEngine'>"}
PASS package_abi_import

• Direct RDMA Transfer Engine tests passed for MUSA VRAM registration,
  sync/batched/async write, sync read, probe/error paths, completion status,
  and payload correctness. The runtime selected RDMA with MC_FORCE_HCA=1.

RDMA runtime evidence:

I0831 03:32:14.629204 transfer_engine_impl.cpp:388] Using RDMA transport (RoCE/iWARP)
{"batch_async_write": "pass", "batch_sync_write": "pass", "bytes": 8388608, "ok": true, "probe_and_error_path": "pass", "protocol": "rdma", "sync_read": "pass", "sync_write": "pass"}
I0831 03:32:32.536921 transfer_engine_impl.cpp:388] Using RDMA transport (RoCE/iWARP)
{"ok": true, "batch_register_memory": "pass", "batch_unregister_memory": "pass", "buffers": 2}

• Direct MUSA IPC/P2P tests passed in both GPU directions for default,
  auto, and transfer_batch: 6/6 cases completed write/readback and
  payload validation. All cases selected the MUSA transport with no RDMA/TCP
  selection.

MUSA IPC/P2P runtime evidence:

I0831 07:08:04.618232 transfer_engine_impl.cpp:379] Using MUSA transport (forced or no HCA detected)
I0831 07:08:04.676371 musa_transport.cpp:195] MusaTransport: copy API auto, transfer-batch min bytes 0
{"bytes": 16777216, "mode": "auto", "ok": true, "payload_ok": true, "protocol": "musa", "role": "initiator", "source_gpu": 0, "target_gpu": 1}
I0831 07:09:11.415910 transfer_engine_impl.cpp:379] Using MUSA transport (forced or no HCA detected)
I0831 07:09:11.419875 musa_transport.cpp:195] MusaTransport: copy API transfer_batch, transfer-batch min bytes 0
{"bytes": 16777216, "mode": "transfer_batch", "ok": true, "payload_ok": true, "protocol": "musa", "role": "initiator", "source_gpu": 1, "target_gpu": 0}

The complete matrix is 2 directions x 3 copy policies; all 6 result records
reported protocol="musa" and payload_ok=true.

• Supplementary TE A1/B/A2 image-level regression stayed within the +/-5%
  threshold: -2.10% for 32 MiB single write and -1.97% for 128 x 256 KiB
  batch write.

Scope

This validates Mooncake MUSA wheel functionality on the tested single-node,
two-GPU S5000 environment with MUSA driver 5.2.0. The MUSA Transfer Engine
implementation comes from the upstream 0.3.13 wheel; this repository carries
the image/build contract and the v0.28 cache-layout compatibility patch.

This PR does not claim vLLM model performance, streaming, FA3, FlashMLA,
cross-node MUSA IPC, long-duration stability, failure recovery, or CUDA-wheel
compatibility. The A1/B/A2 result is an image-level TE regression check, not
strict wheel-only performance attribution.