[Feature][MRV2] GQA C8 modelrunnerv2 adaptation (vllm-project#12981)

### What this PR does / why we need it?
This PR fixes a bug in `modelrunnerv1` and `modelrunnerv2` caused by
inconsistent `attn_state` values ​​during graph mode compilation.

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?

- vLLM version: v0.25.1
- vLLM main:
vllm-project/vllm@fe784ff

---------

Signed-off-by: pichangping <1337510399@qq.com>

Author: pichangping

vllm_ascend/attention/attention_v1.py

@@ -1714,11 +1714,14 @@ def forward(
                 attn_output = self._forward_encoder_attention(query, key, value, attn_metadata, output)
                 output[:num_tokens] = attn_output[:num_tokens]
                 return output
+
+            # When `modelrunnerv2` compiles the graph, the value of `attn_metadata.attn_state` is `None`;
+            # therefore, the graph-mode condition needs to be evaluated earlier.
+            if _EXTRA_CTX.capturing:
+                attn_output, num_tokens = self.full_graph_fia(query, key, value, attn_metadata, output, layer)
+                output[:num_tokens] = attn_output[:num_tokens]
+                return output
             if attn_metadata.attn_state == AscendAttentionState.DecodeOnly:
-                if _EXTRA_CTX.capturing:
-                    attn_output, num_tokens = self.full_graph_fia(query, key, value, attn_metadata, output, layer)
-                    output[:num_tokens] = attn_output[:num_tokens]
-                    return output
                 return self._forward_c8_decode(query, attn_metadata, output, layer)
             elif attn_metadata.attn_state == AscendAttentionState.ChunkedPrefill:
                 return self._forward_c8_chunked_prefill(query, float_key, float_value, attn_metadata, output, layer)