Match Qwen head-chunked attention execution

docs/issue113-head-chunk/checkpoint.log

@@ -0,0 +1,11 @@
+running 1 test
+test engine::metal::qwen_mtplx::session_checkpoint::qwen_checkpoint_roundtrip_and_failed_restore_preserve_live_state ... ds4: Metal device Apple M5 Max, 128.00 GiB RAM
+ds4: Metal 4 tensor API enabled for Tensor kernels
+ds4: drift-patch flags hc_stable=on norm_unify=on kv_raw_f32=off rope_exp2_log2=off math_safe=off tensor_matmul=on
+ds4: Qwen reference kernels: Powered by MTPLX — https://github.com/youssofal/mtplx
+#113 SDPA policy: legacy/current, both directions, AR/MTP, 1/3/4-token frontiers; exact compatible restore and incompatible rejection before tensor reads
+Qwen sorted-gather checkpoint policy: old short state exact, possibly affected long state rejected before restore
+Qwen checkpoint: exact roundtrip, cancellation/corruption controls, old MTP rejected and old AR restored
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 548 filtered out; finished in 0.13s