【main】SP For Qwen3 MoE (#2209)

### What this PR does / why we need it?
Qwen3 MoE supports SP. In scenarios like AlltoAll, AlltoAllv, and MC2,
replacing AllReduce with Reduce-Scatter and AllGather achieves
computational benefits in norm operations while saving one AllGather
communication. This feature is enabled during the P-phase and delivers
notable gains in long-sequence scenarios (e.g., 16k–25k), with
performance improvements reaching 5%–10%.
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
``` 
compilation_config={
    "pass_config":{
        "enable_sequence_parallelism": True
    }
},
enable_expert_parallel=True,
```

- vLLM version: v0.10.0
- vLLM main:
9edd1db02b

---------

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

vllm_ascend/ops/fused_moe.py

@@ -47,6 +47,7 @@ from vllm_ascend.distributed.parallel_state import get_mc2_group
 from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
 from vllm_ascend.ops.moe_dispatcher.token_dispatcher import (
     MoEAlltoAllSeqOverLapDispatcher, MoEDispatcherConfig)
+from vllm_ascend.ops.sequence_parallel import MetadataForPadding
 from vllm_ascend.torchair.utils import npu_stream_switch, npu_wait_tensor
 from vllm_ascend.utils import (AscendSocVersion, dispose_tensor,
                                get_all_reduce_merge_state,
@@ -1347,7 +1348,8 @@ class AscendFusedMoE(FusedMoE):
                 top_k: Optional[int] = None,
                 shared_experts: Optional[Any] = None,
                 gate=None,
-                replace_allreduce: bool = False):
+                replace_allreduce: bool = False,
+                _metadata_for_padding: Optional[MetadataForPadding] = None):
 
         assert self.quant_method is not None
 
@@ -1381,7 +1383,17 @@ class AscendFusedMoE(FusedMoE):
                 # When all_reduce_merge is in progress, shared_experts does not do all_reduce in mlp, but waits until shared_experts+router_experts are completed before doing all_reduce
                 shared_hidden_states = shared_experts(hidden_states)
 
+        mc2_mask = forward_context.mc2_mask
+
+        enable_sp = _metadata_for_padding is not None and _metadata_for_padding.not_dummy_and_is_prefill
         tp_size = get_tensor_model_parallel_world_size()
+        if enable_sp:
+            tp_rank = get_tensor_model_parallel_rank()
+            mc2_mask_sp = _metadata_for_padding.mc2_mask if _metadata_for_padding is not None else forward_context.mc2_mask
+            chunk_mc2_mask = torch.tensor_split(mc2_mask_sp, tp_size, dim=0)
+            mc2_mask = chunk_mc2_mask[tp_rank]
+            replace_allreduce = True
+
         if (fused_moe_state not in [
                 FusedMoEState.AllGather, FusedMoEState.AllGatherEP,
                 FusedMoEState.NaiveMulticast