Add support for topk metadata transferring for PD (#10616)

Signed-off-by: Shangming Cai <csmthu@gmail.com>

python/sglang/srt/managers/schedule_batch.py

@@ -607,6 +607,8 @@ class Req:
             ) = None
         self.hidden_states: List[List[float]] = []
         self.hidden_states_tensor = None  # Note: use tensor instead of list to transfer hidden_states when PD + MTP
+        self.output_topk_p = None
+        self.output_topk_index = None
 
         # Embedding (return values)
         self.embedding = None