[FEAT] Refactor spec decode to support efficient padded speculation (#3528)

### What this PR does / why we need it?
1. Refactor the file `mtp_proposer.py`, splits torchair related codes
into `mtp_torchair_proposer.py`
2. According to https://github.com/vllm-project/vllm/pull/24539,
implements padded speculative decoding as described in
https://github.com/vllm-project/vllm/issues/21984.
### Does this PR introduce _any_ user-facing change?
User can use `disable_padded_drafter_batch` to disable/enable padded
speculation, default is `False`.
offline example:
```
speculative_config={"method": "deepseek_mtp", "num_speculative_tokens": 1, "disable_padded_drafter_batch": False}
```

### How was this patch tested?

- [x] egaer with pad/unpad:
- [x] aclgraph with pad/unpad
- [x] torchair with pad/unpad

performance test of deepseek-r1 with tp16、dp1
aclgraph with pad ITL: 168ms
aclgraph with unpad ITL: 169ms
original: 178ms


- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19

---------

Signed-off-by: xuyexiong <xuyexiong@huawei.com>

vllm_ascend/torchair/torchair_model_runner.py

@@ -34,6 +34,7 @@ from vllm.logger import logger
 import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.platform import NPUPlatform
+from vllm_ascend.spec_decode import get_spec_decode_method
 from vllm_ascend.torchair.utils import (
     TORCHAIR_CACHE_DIR, TorchairCommonAttentionMetadata,
     check_torchair_cache_exist, converting_weight_acl_format,
@@ -83,6 +84,20 @@ class NPUTorchairModelRunner(NPUModelRunner):
 
         self._check_batch_sizes_consistency()
 
+    def _set_up_drafter(self):
+        super()._set_up_drafter()
+        if self.speculative_config:
+            # Torchair do not support disable_padded_drafter_batch
+            # Enforce to disable this feature
+            self.speculative_config.disable_padded_drafter_batch = True
+
+    def _get_drafter(self):
+        return get_spec_decode_method(self.speculative_config.method,
+                                      self.vllm_config,
+                                      self.device,
+                                      self,
+                                      is_torchair_graph=True)
+
     def _may_pad_kv_consumer_num_seq(self):
         # pd disaggregation scenario need redundant_batch_sizes to avoid each batch's seq_len exceed 16 tokens
         # self.max_num_reqs here is greater than the actual maximum request number