[MTP][Bugfix] Fix GLM5-W8A8 precision issues caused by rotary quant MTP weights (#7139)

### What this PR does / why we need it?
When GLM5 target model uses rotary quant, the final hidden states passes
to MTP need to do an extra rotary.

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Wangbingjie <wangbj1207@126.com>
Signed-off-by: wangbj127 <256472688+wangbj127@users.noreply.github.com>

vllm_ascend/patch/__init__.py

@@ -452,3 +452,35 @@
 #       https://github.com/vllm-project/vllm/pull/34880
 #    Future Plan:
 #       Remove this patch when vLLM merges the PR.
+#
+# ** 21. File: worker/patch_deepseek_mtp.py**
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+#   1. `vllm.model_executor.models.deepseek_v2.get_spec_layer_idx_from_weight_name` and
+#      `vllm.model_executor.models.deepseek_mtp.get_spec_layer_idx_from_weight_name`
+#    Why:
+#       When GLM5 uses rotary quant in vllm-ascend, the MTP layer needs to load an extra weight
+#       named `rot.weight`.
+#    How：
+#       If weight name starts with `rot`, return `layer_id + i` like other tensors in MTP layer.
+#    Related PR (if no, explain why):
+#       Rotary quant is a unique feature of vllm-ascend.
+#    Future Plan:
+#       Remove this patch when vllm supports rotary quant or pluggable `MultiTokenPredictorLayer`.
+#   2. `vllm.model_executor.models.deepseek_mtp.DeepSeekMultiTokenPredictorLayer`
+#    Why:
+#       When GLM5 uses rotary quant in vllm-ascend, the `previous_hidden_states` does not .
+#    How：
+#       If the target model uses rotary quant, a new linear operation is added before `ehnorm`.
+#    Related PR (if no, explain why):
+#       Rotary quant is a unique feature of vllm-ascend.
+#    Future Plan:
+#       Remove this patch when vllm supports rotary quant or pluggable `MultiTokenPredictorLayer`.
+#   3. `vllm.model_executor.models.deepseek_mtp.DeepSeekMTP._rewrite_spec_layer_name`
+#    Why:
+#       Rename `rot.weight` to match the format of weights in `DeepSeekMTP`.
+#    How：
+#       If the weight name is `rot`, rename it to `model.layers.{spec_layer}.rot.weight`.
+#    Related PR (if no, explain why):
+#       Rotary quant is a unique feature of vllm-ascend.
+#    Future Plan:
+#       Remove this patch when vllm supports rotary quant or pluggable `MultiTokenPredictorLayer`.