Unpack NMS inputs into bbox, scores and class ids (apache#7257)

commit fe8fda81774c2e1a4d434179f62e3a299e084cb7
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Wed Dec 30 20:31:29 2020 +0900

    fix write by a single thread

commit 0c21e36
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Tue Dec 29 04:32:18 2020 +0900

    minor improvement when topk is available

commit 68c6866
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Tue Dec 29 04:10:24 2020 +0900

    finish concat output

commit 37d7a19
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Tue Dec 29 03:59:28 2020 +0900

    fixed topk handling

commit 1913f97
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Mon Dec 28 21:34:24 2020 +0900

    more refactoring

commit 70c65f0
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Mon Dec 28 21:27:15 2020 +0900

    unpack input data

commit 3a27397
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Mon Dec 28 21:22:16 2020 +0900

    slight change to initialization

commit 9b42008
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Mon Dec 28 19:50:36 2020 +0900

    add some comments, remove check the check on negative class id

commit 0aa375d
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Mon Dec 28 19:39:49 2020 +0900

    leave a TODO on write by only one thread

commit d75ee0a
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Mon Dec 28 19:13:04 2020 +0900

    temp disable write by only thread 0

commit 20b5630
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Sat Dec 26 10:06:43 2020 +0900

    use one block two avoid global sync issue

commit dd1e230
Author: Masahiro Masuda <masahi129@gmail.com>
Date:   Sat Dec 26 07:59:19 2020 +0900

    make NMS inner loop parallel
fix write by a single thread