[Feature] Support kv nz feature for DeepSeek decode node in disagg-prefill scenario (#3072)

By converting the KV cache from ND to NZ format when the decode node
receives it, this PR ensures that the KV NZ feature works correctly
during the decoding phase in disagg-prefill scenario.

- vLLM version: v0.11.0
- vLLM main:
83f478bb19

---------

Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
Co-authored-by: ghphotoframe <854746559@qq.com>
Co-authored-by: alex101-ops <alex1015718386@gmail.com>

tests/ut/test_ascend_config.py

@@ -39,6 +39,7 @@ class TestAscendConfig(TestBase):
         ascend_config = init_ascend_config(test_vllm_config)
         self.assertIsNone(ascend_config.expert_map_path)
         self.assertFalse(ascend_config.multistream_overlap_shared_expert)
+        self.assertFalse(ascend_config.enable_kv_nz)
 
         ascend_compilation_config = ascend_config.ascend_compilation_config
         self.assertTrue(ascend_compilation_config.fuse_norm_quant)
@@ -53,6 +54,7 @@ class TestAscendConfig(TestBase):
             "multistream_overlap_shared_expert": True,
             "expert_map_path": "test_expert_map_path",
             "refresh": True,
+            "enable_kv_nz": False
         }
         ascend_config = init_ascend_config(test_vllm_config)
         self.assertEqual(ascend_config.expert_map_path, "test_expert_map_path")
@@ -61,6 +63,7 @@ class TestAscendConfig(TestBase):
 
         ascend_compilation_config = ascend_config.ascend_compilation_config
         self.assertFalse(ascend_compilation_config.fuse_norm_quant)
+        self.assertFalse(ascend_config.enable_kv_nz)
 
     @_clean_up_ascend_config
     def test_init_ascend_config_enable_npugraph_ex(self):