chore(core): switch backend to spawn

Remove the "forced lower case" when reading VOC class names.

fix(demo): update yolox.cpp in ncnn

fix(demo): fix num_class value in TRT inference

chore(demo): add size adaptive blobFromImage function in TRT

chore(examples): update import in exps

remove redundant code in EMA

* fix typo "wheather" -> "whether" in comment in YOLOXHead Class

Add MegEngine implementation link

Skip to save unused tensor in dynamic_k_matching function for saving memory usage

* add train custom dataset code
* update custom docs
* fix(evaluator): fix n_samples=0 bug for the len of dataloader is 1

Default needs to be set to False, for the option to have any effect at all

fix(data): fix bug in  mixup 

fix(tools): fix bug of mm training

* fix RAM leak error

* fix another memory leak

* update boxes.py

* make flake8 happy

fix(docs): fix typo in ncnn README

* fix bug in demo_postprocessing for non-square input image

fix(docs): fix typo and highlight keyword

This reverts commit 9ac889 to fix training bug

chore(tools): fix potential bug  and update docs 

* fix openvino python demo

* change to subprocess beckend

* fix eval bug

* Update launch.py

make flake8 happy
Co-authored-by: NFeng Wang <wangfeng19950315@163.com>

chore(core): update launch backend to subprocess

chore(demo): add MegEngine benchmark

chore(demo): update megengine deployment