perf: write ref images + masks as QFRAW raw format (skip BMP/PNG)

Pair commit to the QuantFunc backend's QFRAW01/QFRAWL1 fast-path readers.
Replace cv2.imwrite(BMP) for ref images and PIL/cv2 PNG for masks with
a 16-byte header + raw uint8 bytes write to /dev/shm. Backend's
ImageUtils::load_image / load_mask detect the magic bytes and skip
cv::imread / PNG decode entirely.

Format:
  Image: "QFRAW01\0" + uint32_le H + uint32_le W + H*W*3 uint8 RGB
  Mask:  "QFRAWL1\0" + uint32_le H + uint32_le W + H*W   uint8 L

Measured: 1376×1024 edit, 1728×2304 ref → ~60 ms saved per call (BMP
encode/decode round-trip removed from the IPC path).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>