Add fp8 qkv_proj_with_rope kernel for CPU in sgl-kernel and add UT (#6493)

test/srt/cpu/test_decode.py

@@ -1,7 +1,7 @@
 import unittest
 
+import sgl_kernel
 import torch
-from sgl_kernel.common_ops import decode_attention_cpu as decode_attention
 from torch.nn.functional import scaled_dot_product_attention
 
 from sglang.test.test_utils import CustomTestCase
@@ -105,7 +105,7 @@ class TestDecodeAttention(CustomTestCase):
         v_buffer = v_buffer.transpose(0, 1).contiguous().transpose(0, 1)
         key = key.transpose(0, 1).contiguous().transpose(0, 1)
         value = value.transpose(0, 1).contiguous().transpose(0, 1)
-        decode_attention(
+        torch.ops.sgl_kernel.decode_attention_cpu(
             q,
             k_buffer,
             v_buffer,