Remove chunked_prefill_for_mla and fix ring_mla bug (#2781)

### What this PR does / why we need it?
Remove chunked prefill for mla branch in mla , and change dtype of
prefill_mask to avoid accuracy problem
### Does this PR introduce _any_ user-facing change?
NO
### How was this patch tested?

- vLLM version: v0.10.2
- vLLM main:
ef7eefe17a

---------

Signed-off-by: SunnyLee219 <3294305115@qq.com>

examples/disaggregated_prefill_v1/README.md

@@ -70,9 +70,7 @@ vllm serve /models/deepseek_r1_w8a8 \
   "kv_port": "20001",
   "engine_id": "0",
   "kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
-  }'  \
-  --additional-config \
-  '{"chunked_prefill_for_mla":true}' 
+  }'
 ```
 
 Run prefill server P2 on second node:
@@ -114,9 +112,7 @@ vllm serve /models/deepseek_r1_w8a8 \
   "kv_port": "20001",
   "engine_id": "0",
   "kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
-  }'  \
-  --additional-config \
-  '{"chunked_prefill_for_mla":true}'
+  }'
 ```
 
 Run decode server d1 on third node: