[Feature] support aclgraph for model runner v2 (#7110)

### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model. 

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

### How was this patch tested?

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

---------

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>

vllm_ascend/worker/v2/input_batch.py

@@ -22,6 +22,8 @@ import numpy as np
 import torch
 from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
 
+from vllm_ascend.attention.attention_v1 import AscendAttentionState
+
 
 class AscendInputBuffers(InputBuffers):
     """Input buffers for Ascend NPUs."""
@@ -37,6 +39,16 @@ class AscendInputBuffers(InputBuffers):
             max_num_tokens,
             device,
         )
+        del self.query_start_loc
+
+        # NOTE: For FULL mode we change +1 to +2 to reserve extra space for padding.
+        # See _pad_query_start_loc_for_fia.
+        self.query_start_loc: torch.Tensor = torch.zeros(
+            max_num_reqs + 2,
+            dtype=torch.int32,
+            device=device,
+        )
+
         # Create seq_lens_cpu and seq_lens_np.
         # npu's attention backend still needs seq_lens on CPU side.
         self.seq_lens_cpu: torch.Tensor = torch.zeros(
@@ -56,6 +68,8 @@ class AscendInputBatch(InputBatch):
     # Create seq_lens_np.
     # npu's attention backend still needs seq_lens on CPU side.
     seq_lens_np: np.ndarray
+    # attn_state is used to build attention metadata.
+    attn_state: AscendAttentionState | None = None
 
     @classmethod
     def make_dummy(
@@ -79,4 +93,11 @@ class AscendInputBatch(InputBatch):
         input_buffers.seq_lens_np[num_reqs:] = 0
         seq_lens_np = input_buffers.seq_lens_np[:num_reqs]
         input_batch.seq_lens_np = seq_lens_np
+        # A dummy run for dp or memory profiling.
+        # When dummy run for dp, num_tokens is set to 1,
+        # so attn_state is set to DecodeOnly.
+        # when dummy run for memory profiling,
+        # attention metadata isn't needed,
+        # we can also set attn_state to AscendAttentionState.DecodeOnly.
+        input_batch.attn_state = AscendAttentionState.DecodeOnly
         return cls(**asdict(input_batch), seq_lens_np=seq_lens_np)