[long_seq_Feat] support chunk prefill (#4158)

### What this PR does / why we need it?
1、qwen GQA attention_v1 optim
2、DeepSeek MLA refactor, all gather q -> all gather kv 
3、modelrunner refactor for chunk prefill, we remove some code not use

- vLLM version: v0.11.0
- vLLM main:
vllm-project/vllm@2918c1b

---------

Signed-off-by: LookAround <[email protected]>
Signed-off-by: Delphine-Nic <[email protected]>
Co-authored-by: Delphine-Nic <[email protected]>

Author: LookAround0301

tests/ut/attention/test_mla_v1.py

@@ -484,9 +484,6 @@ def test_compute_prefill_context(self, mock_ring, mock_load):
         chunk_ctx.chunk_seq_lens = [torch.tensor([8])]
         chunk_ctx.chunk_seq_lens_npu = [torch.tensor([8])]
         chunk_ctx.starts = [torch.tensor([0])]
-        chunk_ctx.max_chunk_num = 1
-        chunk_ctx.mask_for_non_zero_chunk = [True]
-        chunk_ctx.local_chunked_kv_lens = [[[[8]]]]
 
         prefill_meta = MagicMock()
         prefill_meta.chunked_context = chunk_ctx

vllm_ascend/attention/attention_v1.py

@@ -20,13 +20,6 @@ class AscendPrefillContextParallelMetadata:
 
     num_computed_tokens_of_pcp_dcp: Optional[list[list[list[int]]]] = None
 
-    local_chunked_kv_lens: Optional[list[Optional[list[Optional[list[Optional[
-        list[int]]]]]]]] = None
-
-    mask_for_non_zero_chunk: Optional[List[bool]] = None
-
-    max_chunk_num: int = 0
-
     q_head_idx_tensor: torch.Tensor = None
 
     q_tail_idx_tensor: torch.Tensor = None
@@ -115,23 +108,6 @@ class AscendCommonAttentionMetadata:
         AscendPrefillContextParallelMetadata] = None
 
 
-def extract_req_dcp_by_chunk_pcp(lst,
-                                 chunk_idx,
-                                 dcp_size,
-                                 pcp_rank,
-                                 fill_value=0):
-    num_reqs = len(lst)
-    results: List[List[int]] = []
-    for i in range(num_reqs):
-        if len(lst[i]) == 0 or chunk_idx >= len(lst[i]):
-            # empty req or this req has no corresponding chunk, fill 0
-            results.append([fill_value] * dcp_size)
-            continue
-        dcp_values = lst[i][chunk_idx][pcp_rank]
-        results.append(dcp_values)
-    return results
-
-
 def filter_chunked_req_indices(
     seq_len: torch.Tensor,
     mask_for_non_zero_chunk: Optional[List[bool]],