perf: reduce NV12 per-frame copies from 5 to 2 via single from_buffer_copy + byref offset

tests/benchmark_nv12_decode.py

@@ -11,7 +11,11 @@ mppvideodec auto-selected) and reports:
   • A/V sync drift (video PTS vs pipeline clock position)
   • from_buffer_copy() time per frame (CPU copy cost)
 
-Run on device:
+Run on device (must use same env as the app):
+    export LD_LIBRARY_PATH=/home/ark/miniconda3/envs/r36s-dlna-browser/lib
+    export GST_PLUGIN_PATH=/usr/lib/aarch64-linux-gnu/gstreamer-1.0
+    export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libgomp.so.1
+    export PYTHONPATH=/home/ark/R36SHack/src
     /home/ark/miniconda3/envs/r36s-dlna-browser/bin/python \\
         /home/ark/R36SHack/tests/benchmark_nv12_decode.py [URL]
 
@@ -186,12 +190,13 @@ def _on_sample(sink) -> Gst.FlowReturn:
 
     if fmt_str == "NV12":
         y_size  = int(info.stride[0]) * int(info.height)
-        uv_size = int(info.stride[1]) * (int(info.height) // 2)
-        # Simulate named c_ubyte array creation (the actual upload path)
+        # Simulate the single from_buffer_copy the app now does:
+        # ONE copy of the full Y+UV buffer, then ctypes.byref for the UV offset.
         t1 = time.monotonic()
-        y_arr  = (ctypes.c_ubyte * len(raw[:y_size])).from_buffer_copy(raw[:y_size])
-        uv_arr = (ctypes.c_ubyte * len(raw[y_size:y_size+uv_size])).from_buffer_copy(raw[y_size:y_size+uv_size])
-        copy_us = (time.monotonic() - t1) * 1e6   # override with NV12-specific copy
+        arr = (ctypes.c_ubyte * len(raw)).from_buffer_copy(raw)
+        y_ptr  = ctypes.cast(arr, ctypes.POINTER(ctypes.c_ubyte))
+        uv_ptr = ctypes.cast(ctypes.byref(arr, y_size), ctypes.POINTER(ctypes.c_ubyte))
+        copy_us = (time.monotonic() - t1) * 1e6
 
     with stats.lock:
         stats.total_frames += 1