[Feat] Supports Aclgraph for bge-m3 (#3171)

### What this PR does / why we need it?
[Feat] Supports Aclgraph for bge-m3

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
```
pytest -s tests/e2e/singlecard/test_embedding.py
pytest -s tests/e2e/singlecard/test_embedding_aclgraph.py
```
to start an online server with bs 10, each batch's seq length=8192, we
set --max-num-batched-tokens=8192*10 to ensure encoder is not chunked:
```
vllm serve /home/data/bge-m3 --max_model_len 1024 --served-model-name "bge-m3" --task embed --host 0.0.0.0 --port 9095 --max-num-batched-tokens 81920 --compilation-config '{"cudagraph_capture_sizes":[8192, 10240, 20480, 40960, 81920]}'
```
For bs10, each batch's seq length=8192, QPS is improved from 85 to 104,
which is a 22% improvement, lots of host bound is reduced.


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

---------

Signed-off-by: xuyexiong <xuyexiong@huawei.com>
Co-authored-by: wangyongjun <1104133197@qq.com>

vllm_ascend/attention/attention_mask.py

@@ -50,6 +50,7 @@ class AttentionMaskBuilder:
         self._seq_len_cached = attn_mask.shape[0]
         self.attn_mask_cache = attn_mask
         self.device = device
+        self.pooling_mask = None
         if torch.version.cann.startswith("8.3"):
             assigned_mask_dim = 2048
             self.chunked_prefill_attn_mask = torch.triu(
@@ -75,6 +76,14 @@ class AttentionMaskBuilder:
         return self.attn_mask_cache[:max_seq_len, :max_seq_len].contiguous(
         ).to(device, non_blocking=True)
 
+    def get_pooling_mask(self, device):
+        if self.pooling_mask is None:
+            # the compressed attention mask for npu_fusion_attention sparse mode 4
+            self.pooling_mask = torch.triu(torch.ones(
+                2048, 2048), diagonal=1).to(torch.bool).to(device,
+                                                           non_blocking=True)
+        return self.pooling_mask
+
     def get_splitfuse_attn_mask(
         self,
         seq_lens: torch.Tensor = None,