fix: fix input shape to [192, 48]

* add more distributed training speed-info

* fix header

turn off the mkldnn and trt in ConvNeXt infer

it will time out when turn on trt or mkldnn

Put the calling of update_res(), stop_after(), freeze_befor() to the construction method of the parent class. And the init_net() support to call stop_after() by config of Arch.

freeze subnet specified with the specified layer name by setting OutputTensor.stop_gradient=True

add the config of orientation model
add the preprocess op RandomRot90 that can rotate the img and return the orientation
add the CustomLabelDataset that support getting label by preprocess
refactor some preprocess ops to support dict parameter and return dict

add shitu pipeline evaluation

Add flexible configuration of PP-ShiTu detection model

Add shitu android demo2

the ips is unstable in the first 50 iter when training lite modle.

update table_attribute demo imags

fix shitu scenarios doc