Some XNOR-net improvements

[AlexeyAB]

Some XNOR-net improvements:

• binary output (bit-1 already packed to uint32_t and transposed)
• fuse conv and shortcut with binary output (currently fused with FP32-output)
• fuse conv and maxpool layers with binary output
• remove bottlenecks (on CPU - Fast NMS, 1st conv-layer to UINT8, last conv-layers to FP16)
• allow network input from GPU-memory and on-GPU resizing and convertin GpuMat to image
• improve speed GEMM-XNOR without AVX (unroll 4x4 or 8x8)
• use GEMM from OpenCV for FP32
• support depthwise/group convolutional for xnor-net
• svr https://www.researchgate.net/profile/Hiroki_Nakahara/publication/323375650_A_Lightweight_YOLOv2_A_Binarized_CNN_with_A_Parallel_Support_Vector_Regression_for_an_FPGA/links/5b59ee76a6fdccf0b2f8efe8/A-Lightweight-YOLOv2-A-Binarized-CNN-with-A-Parallel-Support-Vector-Regression-for-an-FPGA.pdf and https://www.researchgate.net/publication/221112464_On-Line_Support_Vector_Machine_Regression/download

[lqian]

great!!!@AlexeyAB