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>

vllm_ascend/ops/moe/moe_mlp.py

@@ -79,8 +79,6 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
 
     is_mc2 = get_forward_context().moe_comm_type == MoECommType.MC2
     if w1_scale_bias is None and is_mc2:
-        if w1_scale.dtype != torch.float32:
-            w1_scale = w1_scale.to(torch.float32)
         if fusion:
             # gmm1: gate_up_proj & act_fn: swiglu
             hidden_states, swiglu_out_scale, _ = torch_npu.npu_grouped_matmul_swiglu_quant(
@@ -90,6 +88,8 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
                 weight_scale=w1_scale,
                 x_scale=pertoken_scale)
         else:
+            if w1_scale.dtype != torch.float32:
+                w1_scale = w1_scale.to(torch.float32)
             # gmm1: gate_up_proj
             hidden_states = torch_npu.npu_grouped_matmul(
                 x=[hidden_states],