[P/D][main]Offline the llmdatadist connector related parts of the code and files. (#4780)

### What this PR does / why we need it?
As support for the mooncake connector is now available, the llmdatadist
connector is no longer being maintained, so the llmdatadist-related
files need to be retired.

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

### How was this patch tested?
By ci

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Signed-off-by: liziyu <liziyu16@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>

vllm_ascend/worker/model_runner_v1.py

@@ -3398,7 +3398,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
         # init kv cache tensors
         kv_cache_raw_tensors: dict[str, Union[torch.Tensor,
                                               Optional[torch.Tensor]]] = {}
-        # llmdatadist need the addr of cache tensor be aligned with 2M
+        # prefill disaggregation need the addr of cache tensor be aligned with 2M
         alignment = 2 * 1024 * 1024
         for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
             # TODO: REFACTOR ME to sharing hybrid cache
@@ -3426,7 +3426,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
                 elif "attn" in layer_name and layer_name not in kv_cache_raw_tensors.keys(
                 ):
                     # NOTE: We need to init k cache tensor (nope cache tensor in mla) and
-                    # v cache tensor (rope cache tensor in mla) separately to support llmdatadist,
+                    # v cache tensor (rope cache tensor in mla) separately to support prefill disaggregation,
                     # as it only support the 0-dim of kv_cache is `num_blocks`.
                     # For deepseek mla, we need to spilt cache tensor accrodding to the nope head dim
                     # and rope head dim.