feat(intel-xpu): enable distributed Wan tensor-parallel inference

[qiuxin2012]

Summary

Enable distributed tensor-parallel inference for Wan models on Intel XPU.

Changes

• Initialize Intel XPU distributed environments with the XCCL backend.
• Bind each process to its LOCAL_RANK.
• Guard unsupported non-Linux distributed initialization based on the PyTorch version.
• Add an opt-in Wan TP weight-loading mode:
  • broadcast remains the default to preserve existing behavior.
  • local lets each rank load and retain its local TP shard, avoiding large XCCL weight broadcasts.
• Fix TP row-parallel bias handling during CPU/XPU state migration and block offload.
• Add a Wan2.2 TI2V TP2 example configuration and launch script for Intel XPU.

Motivation

Large tensor broadcasts during Wan TP model initialization can crash in the Intel XPU/XCCL stack. Local TP weight loading avoids this initialization-time broadcast while keeping
the existing loading behavior unchanged by default.

Usage

bash scripts/platforms/intel_xpu/dist_infer/run_wan22_ti2v_t2v_tp.sh

The example uses two XPU devices and enables local TP weight loading with:

{
 "parallel": {
   "seq_p_size": 1,
   "tensor_p_size": 2,
   "tp_load_mode": "local",
   "cfg_p_size": 1
 }
}

Validation

• Python compilation checks passed.
• Shell syntax validation passed.
• JSON validation passed.
• Successfully ran Wan2.2 TI2V inference with TP2 on two Intel XPU B60s.