Created by: luotao1

speed up the im2col function：

• decreasing the calculation: for example, int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1]; needs channels_col* col_height* col_width * (2 mul + 1sum + 1 minus), now, it needs channels_col* col_height* col_width * 1sum + channels_col * (1sum + 2minus)
• use flag = im_row_idx < 0 || im_row_idx >= im_height || im_col_idx < 0 || im_col_idx >= im_width; to increase the hit of if.

#10685 (closed) OCR CRNN_CTC model:

• the conv2d time: 4.83ms(before) -> 4.41ms(after), speedup: +8.7%
• whole inference time: 58.6ms(before) -> 53.2ms(after), speedup: 9.2%