[npu] support triangle attention for llama (#5130)

* update fused attn

* update spda

* tri attn

* update triangle

* import

* fix

* fix

colossalai/kernel/cuda_native/mha/flash_attn_2.py

@@ -29,7 +29,6 @@ except ImportError:
     HAS_FLASH_ATTN = False
 
 if HAS_FLASH_ATTN:
-    pass
 
     from .utils import SeqLenInfo
 

colossalai/kernel/cuda_native/mha/mha.py

@@ -44,6 +44,7 @@ class ColoAttention(torch.nn.Module):
         key: torch.Tensor,
         value: torch.Tensor,
         attn_mask: Optional[torch.Tensor] = None,
+        origin_attn_mask: Optional[torch.Tensor] = None,
         attn_mask_type: Optional[AttnMaskType] = None,
         bias: Optional[torch.Tensor] = None,
     ):