Optimize copy_kv_cache for spec decoding (#11126)

Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

python/sglang/srt/model_executor/model_runner.py

@@ -1672,6 +1672,9 @@ class ModelRunner:
                     enable_memory_saver=self.server_args.enable_memory_saver,
                     start_layer=self.start_layer,
                     end_layer=self.end_layer,
+                    enable_kv_cache_copy=(
+                        self.server_args.speculative_algorithm is not None
+                    ),
                 )
 
         # Initialize token_to_kv_pool_allocator