- Online mode: decouple ZMQ receive (producer) and detection (consumer)
  into separate threads to eliminate accumulating latency
- Add DetectorCache to pre-compute ArUco params/dictionaries/object
  points once, reused every frame
- Single-camera fast path: extract only the needed quadrant instead of
  splitting all 4, reducing image copy by 75%
- Add --debug flag for verbose diagnostic logging
- Add ARM (AArch64) cross-compile support via toolchain-aarch64.cmake
- CMakeLists.txt auto-detects x86/arm from CMAKE_SYSTEM_PROCESSOR,
  selects matching thirdparty protobuf/zmq libraries
- Fix kMaxMessageBytes 256MB->64MB to avoid OOM on ARM devices
- Add .gitignore for build artifacts, deploy dirs, and data files
- Update README.md with new architecture, parameters, and ARM usage
- Add run.sh online_detect for ARM device deployment