[Feat][SP] Suport SP for VL MoE models (#7044)

### What this PR does / why we need it?

2nd PR for https://github.com/vllm-project/vllm-ascend/issues/5712,
extend SP to VL MoE models.


### Does this PR introduce _any_ user-facing change?
remove `sp_threshold` in additional config and reuse `sp_min_token_num`
from vLLM.


### How was this patch tested?
- Model: Qwen3-VL-30B-A3B, 
- TP4 DP2
- 100 reqs
- max concurrency 1

| Seq length | Mean TTFT (ms) main | Mean TTFT (ms) this PR |
|------------|---------------------|------------------------|
| 4k         | 429.40               | 323.3                  |
| 16k        | 1297.01              | 911.74                |

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e

---------

Signed-off-by: realliujiaxu <realliujiaxu@163.com>

vllm_ascend/platform.py

@@ -156,6 +156,16 @@ class NPUPlatform(Platform):
     def get_device_capability(cls, device_id: int = 0):
         return None
 
+    @classmethod
+    def apply_config_platform_defaults(cls, vllm_config: VllmConfig) -> None:
+        """Apply Ascend-specific defaults. Set sp_min_token_num=1 when enable_sp and not set."""
+        pass_config = vllm_config.compilation_config.pass_config
+        if pass_config.enable_sp and pass_config.sp_min_token_num is None:
+            from vllm_ascend.compilation.passes.sequence_parallelism import get_sp_min_token_num
+
+            pass_config.sp_min_token_num = get_sp_min_token_num(vllm_config)
+            logger.info(f"set sp_min_token_num to {pass_config.sp_min_token_num}")
+
     @classmethod
     def get_device_name(cls, device_id: int = 0) -> str:
         return torch.npu.get_device_name(device_id)
@@ -198,6 +208,7 @@ class NPUPlatform(Platform):
 
         # initialize ascend config from vllm additional_config
         cls._fix_incompatible_config(vllm_config)
+
         ascend_config = init_ascend_config(vllm_config)
 
         if vllm_config.kv_transfer_config is not None:
@@ -218,6 +229,7 @@ class NPUPlatform(Platform):
                 if not isinstance(ascend_compilation_config, dict)
                 else ascend_compilation_config
             )
+
         ascend_config.update_compile_ranges_split_points()
 
         if model_config and hasattr(model_config.hf_text_config, "index_topk"):
@@ -363,7 +375,8 @@ class NPUPlatform(Platform):
 
         if parallel_config and parallel_config.worker_cls == "auto":
             # TODO: this is a tricky way to disable `use_sequence_parallel_moe` in vllm.
-            parallel_config.all2all_backend = "flashinfer_all2allv"
+            if not vllm_config.compilation_config.pass_config.enable_sp:
+                parallel_config.all2all_backend = "flashinfer_all2allv"
             if is_310p():
                 parallel_config.worker_cls = "vllm_ascend._310p.worker_310p.NPUWorker310"
             elif ascend_config.xlite_graph_config.enabled:
@@ -805,3 +818,7 @@ class NPUPlatform(Platform):
                     "ignored on Ascend. Resetting to default (32)."
                 )
                 att_config.flash_attn_max_num_splits_for_cuda_graph = 32
+
+    @classmethod
+    def use_custom_op_collectives(cls) -> bool:
+        return True