[Feat][Graph] Support FULL_DECODE_ONLY mode for GQA/MHA models (#2128)

Note: This depends on [vLLM
#25161](https://github.com/vllm-project/vllm/pull/25161) and the
torch\_npu release from September 30.

### What this PR does / why we need it?
This pull request adds `FULL_DECODE_ONLY` mode for GQA/MHA models (MLA
models like DeepSeek V3/R1 are not included). Key improvements include:

* **Reduced dispatch latency:** By replaying the entire model execution
graph at once, we cut overhead compared with multiple smaller replays.
* **Stabilized multi-device performance:** Captureing the whole model as
one static graph also mitigates the dispatch fluctuations across
devices.
* **Stream/resource savings:** Consolidating graph captures frees up
streams, allowing more graphs to be captured.

**Known issues:**

1. `_npu_paged_attention` currently manages its own workspace in
`torch_npu`, which can deadlock when synchronizing during graph replay —
we’re working on a fix.

There may be other corner cases. This PR is the first in a planned
series; we’ll continue to iterate and address remaining issues in
follow-ups.

This is essentially a port of #1503 and #1677, but includes two major
changes:

1. Let `graph_dispatcher` decide the graph mode instead of hard-coding
it in the backend, which decouples Full Graph and Piecewise Graph and
could make it possible to remove dynamo.
2. Adapt to the new `attn_group` logic, but leave a small hack in
`update_graph_params`; multi-attention models may or may not be fully
supported yet.

### Does this PR introduce _any_ user-facing change?
```python
compilation_config={
    "cudagraph_mode": "FULL_DECODE_ONLY",
},
```

### How was this patch tested?
Tests included.


- vLLM version: v0.10.2
- vLLM main:
9607d5eb44

---------

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

tests/ut/attention/test_attention_v1.py

@@ -101,7 +101,7 @@ class TestAscendAttentionMetadataBuilder(TestBase):
             max_query_len=5,
             decode_token_per_req=torch.tensor([1, 1]),
             block_table_tensor=torch.zeros((10, 10)),
-            slot_mapping_cpu=torch.tensor(range(20)),
+            slot_mapping=torch.tensor(range(20)),
             actual_seq_lengths_q=torch.tensor([0, 1]),
             positions=torch.tensor([10, 10]),
             attn_mask=torch.ones((10, 10)),
@@ -134,7 +134,7 @@ class TestAscendAttentionMetadataBuilder(TestBase):
             max_query_len=6,
             decode_token_per_req=torch.tensor([1, 1, 1]),
             block_table_tensor=torch.zeros((10, 10)),
-            slot_mapping_cpu=torch.tensor(range(20)),
+            slot_mapping=torch.tensor(range(20)),
             actual_seq_lengths_q=torch.tensor([0, 1, 2]),
             positions=torch.tensor([10, 10]),
             attn_mask=torch.ones((15, 15)),
@@ -165,7 +165,7 @@ class TestAscendAttentionMetadataBuilder(TestBase):
             max_query_len=6,
             decode_token_per_req=torch.tensor([1, 1, 1]),
             block_table_tensor=torch.zeros((10, 10)),
-            slot_mapping_cpu=torch.tensor(range(20)),
+            slot_mapping=torch.tensor(range(20)),
             actual_seq_lengths_q=torch.tensor([0, 1, 2]),
             positions=torch.tensor([10, 10]),
             attn_mask=torch.ones((15, 15)),
@@ -378,10 +378,12 @@ class TestAscendAttentionBackendImpl(TestBase):
         mock_flash_attention_qlens.assert_called_once()
         assert output.shape == (10, 8 * 64)
 
+    @patch('vllm_ascend.attention.attention_v1.get_forward_context')
     @patch('torch_npu._npu_reshape_and_cache')
     @patch('torch_npu._npu_paged_attention')
     def test_forward_decode_only(self, mock_paged_attention,
-                                 mock_npu_reshape_and_cache):
+                                 mock_npu_reshape_and_cache,
+                                 mock_get_forward_context):
         """Test forward pass in DecodeOnly state"""
         query = torch.randn(10, 8 * 64)
         key = torch.randn(10, 8 * 64)
@@ -395,6 +397,8 @@ class TestAscendAttentionBackendImpl(TestBase):
         metadata.slot_mapping = torch.zeros(10, dtype=torch.long)
         layer = self.layer_no_quant
 
+        mock_get_forward_context.return_value = MagicMock(capturing=False)
+
         output = self.impl.forward(layer,
                                    query,
                                    key,
@@ -435,12 +439,13 @@ class TestAscendAttentionBackendImpl(TestBase):
         mock_fused_infer_attention_score.assert_called_once()
         assert output.shape == (10, 8 * 64)
 
+    @patch('vllm_ascend.attention.attention_v1.get_forward_context')
     @patch('torch_npu._npu_reshape_and_cache')
     @patch('torch_npu._npu_paged_attention')
     @patch('torch_npu.npu_fused_infer_attention_score')
     def test_forward_decode_only_swa_seq_len_mismatch(
             self, mock_fused_infer_attention_score, mock_paged_attention,
-            mock_npu_reshape_and_cache):
+            mock_npu_reshape_and_cache, mock_get_forward_context):
         """Test forward pass in DecodeOnly state when seq)len_mismatch"""
         query = torch.randn(10, 8 * 64)
         key = torch.randn(10, 8 * 64)
@@ -457,6 +462,8 @@ class TestAscendAttentionBackendImpl(TestBase):
         mock_fused_infer_attention_score.return_value = (torch.ones(10, 8,
                                                                     64), 1)
 
+        mock_get_forward_context.return_value = MagicMock(capturing=False)
+
         output = self.impl_swa.forward(self.layer_no_quant,
                                        query,
                                        key,