Reimplement Sparse R-CNN (#4219)

* draft version train and test

* remove get_bbox in sparse_head

* refactor SparseRCNN model, align inference eval mAP with official repo

* add focal_loss in hungarian_assigner.py

* update hungarian_assigner.py

* update detr config

* update doc and codes

* fix unitest

* fix unitest

* add embedding_rpn and traing code

* modify code format

* fix proposalist in simpletest

* add focal loss traing

* fix acc

* fix reduce mean

* add embedding rpn doc and add show initproposals

* add docstr

* fix backbone lr

* fix typo and change detach_proposal_list to proposal_list

* remove debug assert

* add readme

* add config for random crop when use 300 proposals

* add aug for 300 proposals

* add 300 proposals results and fix docstr

* add 1x results

* fix lint

* fix unitest

* fix docstr of embeddingrpn

* fix according to the comments

* fix get_bbox and a bug of dynamic conv

* support cpu focal loss and fix docstr

* fix docstr

* fix unitest

* fix bug of  init_weight

* fix to device in assign

* fix all commemt

* fix test_config

* remove dim 1 in forward

* fix shape in comment

* fix the case that train_cfg is None

* fix shape of docstr

* fix docstr

* fix typo

* simplify test

* fix docstr

* move train_cfg and test_cfg to model as

* fix unitest

* add link of ckpt and log

* fix 1x results

Co-authored-by: tianyuandu <tianyuandu@gmail.com>
Co-authored-by: fangkairen <fangkairen@sensetime.com>