Add XPU SYCL GGUF ops for dequantize and A8 matmul

[Copilot]

This PR adds the minimum GGUF/XPU operator surface needed to cover ggml_dequantize, ggml_mul_mat_vec_a8, and ggml_mul_mat_a8. The implementation now provides these operators through the native _C extension using a SYCL kernel path aligned with the existing XPU kernel style in this repository.

• What changed

  • Added GGUF operator registrations for:
    • ggml_dequantize
    • ggml_mul_mat_vec_a8
    • ggml_mul_mat_a8
  • Exported the new interfaces from vllm_xpu_kernels and exposed them through the test op registry.
  • Wired the new implementation into the _C extension build and torch bindings.

• Implementation

  • Replaced the earlier Python compatibility fallback with a native C++/SYCL implementation in csrc/gguf.cpp.
  • Added _C extension declarations and registrations in:
    • csrc/ops.h
    • csrc/torch_bindings.cpp
    • CMakeLists.txt
  • Kept vllm_xpu_kernels/gguf_interface.py as a thin Python wrapper over the registered native ops.
  • The current native GGUF path supports:
    • Q4_0
    • Q8_0

• Dependency surface

  • Removed the previous Python gguf runtime dependency from requirements.txt, since the operator path is now implemented natively in the extension.

• Focused coverage

  • Updated targeted tests to validate:
    • packed Q4_0 / Q8_0 reference construction
    • dequantization behavior
    • vector-style A8 matmul path
    • matrix-style A8 matmul path
  • Tests skip XPU-only execution when the XPU device or built extension is unavailable.

Example usage:

import torch
from vllm_xpu_kernels import (
    ggml_dequantize,
    ggml_mul_mat_a8,
    ggml_mul_mat_vec_a8,
)

weight = torch.tensor(..., device="xpu", dtype=torch.uint8)
x = torch.randn(4, 4096, device="xpu", dtype=torch.float16)

deq = ggml_dequantize(weight, quant_type, m, n, torch.float16)
y_vec = ggml_mul_mat_vec_a8(weight, x[:1], quant_type, row_count)
y_mat = ggml_mul_mat_a8(weight, x, quant_type, row_count)

---

📍 Connect Copilot coding agent with Jira, Azure Boards or Linear to delegate work to Copilot in one click without leaving your project management tool.

[zhenwei-intel]

@copilot 参考fp8_quant.cpp的实现，要用sycl kernel实现

[Copilot]

@copilot 参考fp8_quant.cpp的实现，要用sycl kernel实现

已改成基于 C++/_C 扩展的 SYCL kernel 实现，并去掉了之前的 Python fallback 路径，见 f1c50ad。当前新增了原生 ggml_dequantize、ggml_mul_mat_vec_a8 和 ggml_mul_mat_a8 的注册与实现，测试也同步调整了。