[Feat]: Add custom lmhead tensor model parallel (#2309)

### What this PR does / why we need it?
This PR introduces LMhead tensor model parallel to achieve decreasing of
memory consumption, and TPOT performance improvement. It support both
eager mode and graph mode.

In deepseek r1 w8a8 PD disagregated Decode instance, using pure DP, with
lmhead_tensor_parallel_size = 8, we have 1 ms TPOT optimization, saved
1.48 GB NPU memory per RANK.

performance data:
<img width="1444" height="438" alt="image"
src="https://github.com/user-attachments/assets/3c5ef0d3-a7c7-46fd-9797-4de728eb0cb0"
/>

### Does this PR introduce _any_ user-facing change?
This PR introduces one new config in `additional_config`.
| Name | Effect | Required | Type | Constraints |
| :---------------------------- |
:--------------------------------------- | :------- | :--- |
:----------------- |
| lmhead_tensor_parallel_size | Split the lm_head matrix along the
column dimension (vocab_size) into lmhead_tensor_parallel_size pieces |
No | int | default value is None, once this value is set, the feature
will be enabled, vocab_size must be divisible by this value. |

example

`--additional_config={"lmhead_tensor_parallel_size": 8}`

### How was this patch tested?


- vLLM version: v0.10.1.1
- vLLM main:
de533ab2a1

---------

Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Co-authored-by: zhangzihang <zzh_201018@outlook.com>

vllm_ascend/worker/mtp_proposer_v1.py

@@ -19,7 +19,7 @@ from vllm_ascend.ascend_forward_context import set_ascend_forward_context
 from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
 from vllm_ascend.models.deepseek_mtp import CustomDeepSeekMTP
 from vllm_ascend.torchair.utils import TorchairCommonAttentionMetadata
-from vllm_ascend.utils import ProfileExecuteDuration
+from vllm_ascend.utils import ProfileExecuteDuration, lmhead_tp_enable
 
 
 class MtpProposer:
@@ -235,8 +235,20 @@ class MtpProposer:
                         previous_hidden_states=self.
                         hidden_states[:num_input_tokens],
                         kv_caches=self.runner.kv_caches[-1:])
+
+        num_indices = last_token_indices.shape[0]
+        if lmhead_tp_enable():
+            if not self.runner.with_prefill:
+                max_num_reqs_across_dp = num_input_tokens
+            else:
+                max_num_reqs_across_dp = self.vllm_config.scheduler_config.max_num_seqs
+            last_token_indices = nn.functional.pad(
+                last_token_indices, (0, max_num_reqs_across_dp - num_indices))
+
         sample_hidden_states = hidden_states[last_token_indices]
         logits = self.model.compute_logits(sample_hidden_states, None)
+        if lmhead_tp_enable() and num_indices < logits.shape[0]:
+            logits = logits[:num_indices]
         draft_token_ids = logits.argmax(dim=-1)
 
         # [batch_size, 1]