Remove qwen3 moe MC2 cumsum & cast (#3126)

What this PR does / why we need it?
The Qwen3 moe MC2 graph currently has two redundant computational
operator implementations. After npu_moe_distribute_dispatch_v2, the
cumsum and cast operations have been added. By using
expert_token_nums_type=0 and not converting weight_scale to float32,
these two operators can be eliminated, thereby improving inference
performance.

Does this PR introduce any user-facing change?
No

How was this patch tested?
No need

vLLM version: v0.10.2
vLLM main:
f225ea7dd9

- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9

---------

Signed-off-by: florenceCH <gaoxiang120@huawei.com>
Co-authored-by: florenceCH <gaoxiang120@huawei.com>

tests/ut/ops/test_token_dispatcher.py

@@ -98,7 +98,7 @@ class TestTokenDispatcherWithMC2(TestBase):
                                                     self.row_idx, expert_map)
             mock_dispatch.assert_called_once()
             self.assertEqual(output["group_list_type"],
-                             1)  # group_list_type == 1
+                             0)  # group_list_type == 0
 
     def test_token_dispatch_with_shared_experts_and_quant(self):
         self.shared_experts = MagicMock()