Match Qwen head-chunked attention execution

docs/issue113-head-chunk/glm.log

@@ -0,0 +1,11 @@
+ds4: GLM MTP draft 5733 at position 12 in 2.3 ms
+verify2 scalar boundary cap=2 cancelled=true passed
+verify2 prefill 0
+verify2 prefill 4
+verify2 prefill 8
+verify2 prefill 12
+ds4: GLM MTP draft 5733 at position 12 in 2.3 ms
+engine-lifetime counters across new and restored chats passed
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 548 filtered out; finished in 11.91s