[hardware] feat: add Intel XPU device support

[kahlun]

Summary

Add Intel XPU (GPU) device detection and backend support to VeOmni, enabling training on Intel Arc/Data Center GPUs.

Changes (4 files, +13 / -3 lines)

File	Change	Bug ID
veomni/utils/device.py	Add IS_XPU_AVAILABLE flag, XPU branch in get_device_type(), get_dist_comm_backend() (xccl), and stream_synchronize()	B1, B2
veomni/ops/kernels/moe/_kernels/utils/device.py	Guard torch.cuda.get_device_capability() for non-CUDA devices	B3
veomni/distributed/moe/moe_layer.py	Guard group_gemm import (Triton CUDA kernels) for CUDA-only	B3
veomni/distributed/torch_parallelize.py	Accept "xpu" as valid init_device in non-FSDP path	B5

Testing

Tested on Intel Arc Pro B70 (Battlemage BMG-G31, 32 GB VRAM):

• 1-GPU: 7/7 pass (import, model build, FSDP2 parallelize, optimizer, fwd+bwd, 3-step train loop, CPU offload round-trip)
• 2-GPU: 8/8 pass (with CCL_ATL_SHM=1)
• Model: Qwen2.5-0.5B (494M params, bf16)

Context

Follow-up to #648 (src_data_rank=None fix for scatter hang). These are the remaining XPU device detection gaps that block VeOmni from running on Intel GPUs.

The corresponding veRL integration PR (1-line EngineRegistry registration) is staged separately.

[gemini-code-assist]

Code Review

This pull request introduces support for XPU devices across various modules, including device detection, communication backends, and synchronization utilities. Feedback indicates that the new guard in get_device_key incorrectly returns "unknown" for XPU devices, bypassing device-specific configurations. Furthermore, the conditional import of group_gemm kernels in moe_layer.py is likely to cause a NameError on XPU systems when those kernels are referenced at runtime.

[gemini-code-assist]

This guard returns "unknown" for XPU devices, which prevents the function from reaching the get_device_name() call at the end of the function. This will cause MoE kernels to use fallback configurations instead of device-specific ones. Since this PR aims to add XPU support, the logic should be updated to allow XPU devices to fall through to the device name detection while still guarding the CUDA-specific get_device_capability() calls to avoid crashes.

[gemini-code-assist]

Guarding the import of group_gemm kernels with torch.cuda.is_available() will cause a NameError at runtime on XPU devices when EPGroupGemm or EPMergedFc1GroupGemm are used, as these classes refer to the imported functions in their methods. If MoE is not yet supported on XPU, it would be better to provide a clear error message (e.g., a RuntimeError in the forward method) or ensure the functions are defined as None and checked before use, rather than allowing a NameError to occur.