Encapsulate the second stage into RoI heads (#1999)

* create roi_head

* create roi_head
mv shared_head into roi_head

* fix conflict

* refactor(cascade rcnn): create cascade head in roi head, simplify cascade rcnn

* type: fix
scope: cascade_head
subject: mv stage_loss_weight from train_cfg to roi_head creation

* type: fix
scope: roi_heads
subject: handle the case when train_cfg is None

* type: fix
scope: roi_heads mask inference
subject: fix test_cfg.rcnn bug to test_cfg

* fix mask eval bug

* refactor(double_head): refactor double head roi head

* fix(double_head): clean code and fix __init__.py

* refactor(mask scoring): create mask scoring roi head

* refactor(htc_head): add htc_head

* fix(htc_head): add htc_head in __init__.py

* fix (htc_head): fix htc_head bugs

* fix (ms_roi_head): rm test_cfg.rcnn to test_cfg

* feature (grid rcnn): grid roi head

* fix (grid_head): fix grid head bug

* reformat and change all cfgs

* reformat (roi_head): reformat __init__.py for isort=4.3.21

* fix roi_head test bug

* fix (carafe): change carafe cfgs to use roi_head

* fix (roi_head): fix missing roi_head refactor

* reformat to pass CI

* test all cfgs

* match keys of configs with previous commit 77d073a

* add and pass unittest for all roi heads

* Refactor (roi_head): extract abstract base class for roi head

* Refactor (roi_head): refactor init functions

* Refactor (roi_head): weight init

* Refactor (roi_head): add _bbox_forwrd & _mask_forward as basic functions

* Fix (grid_roi_head): fix bug in bbox_forward_train

* Refactor (roi_head): change to use img.device in forward_dummy)

* Refactor (roi_heads): simplify init functions and _mask_forward

* Fix (cascade_roi_head): fix init bug of cascade roi_head

* Refactor (roi_head): use dict as outputs of _bbox_forward and _mask_forward

* Refactor (test_config): scan valid configs rather than list them all