[Feat] Shared expert dp for deepseek and deepseek_mtp (#3495)

### What this PR does / why we need it?
shared expert dp for deepseek and deepseek_mtp, could be combined with
sp to improve performance.

### How was this patch tested?

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>

tests/ut/models/test_deepseek_mtp.py

@@ -57,6 +57,8 @@ class TestCustomDeepSeekMultiTokenPredictorLayer(PytestBase):
                             'eh_proj',
                             return_value=torch.randn(2, 3, 768))
         mocker.patch("torch.cat", return_value=torch.randn(2, 3, 768))
+        mocker.patch("torch.ops.vllm.maybe_all_gather_and_maybe_unpad",
+                     lambda x, label: x)
         mtp_layer.mtp_block.return_value = (torch.randn(2, 3, 768),
                                             torch.randn(2, 3, 768))
 
@@ -182,6 +184,8 @@ class TestCustomDeepSeekMTP(PytestBase):
         assert isinstance(mtp, CustomDeepSeekMTP)
 
     def test_forward(self, mocker: MockerFixture, setup_mtp):
+        mocker.patch("torch.ops.vllm.maybe_all_gather_and_maybe_unpad",
+                     lambda x, label: x)
         input_ids = torch.tensor([[1, 2, 3]])
         positions = torch.tensor([[0, 1, 2]])
         kv_caches = [torch.tensor([[0.1, 0.2, 0.3]])]