fix(RCLane): stabilize lane decode and ONNX inference

perf(RCLane): bound and cache F1 lane decoding

perf(RCLane): saturate dual H200 train and eval

perf(RCLane): speed up CARLA F1 eval ~14x (downscaled IoU + cached masks)

feat(RCLane): add resume, F1 eval, full logging & per-epoch HF upload for CARLA

feat(RCLane): add PyTorch port of RCLane