[1.7.x] Fix memory leaks in Gluon

[leezu]

Backport #18328

[mxnet-bot]

Hey @leezu , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [unix-cpu, unix-gpu, windows-gpu, sanity, clang, centos-gpu, miscellaneous, edge, website, centos-cpu, windows-cpu]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[leezu]

@mxnet-bot run ci [all]

[ciyongch]

@TaoLv @pengzhao-intel please take a review and help to merge it :)

[TaoLv]

Please also make sure it's fixed on branch v1.x.

[ciyongch]

@TaoLv here's the PR for v1.x branch: #18359

[leezu]

@zhreshold should we revert this PR on the 1.7 and 1.x branch as weakref doesn't work with deepcopy?

https://github.com/dmlc/gluon-cv/blob/428ee05d7ae4f2955ef00380a1b324b05e6bc80f/gluoncv/data/transforms/presets/yolo.py#L144

(Though I'm surprised that copy.deepcopy worked in the first place; for example, using deepcopy probably leads to double free and/or use of invalid memory as you duplicate the ndarray handles of the parameters)

[zhreshold]

@leezu Fix to mem leak is nice to have and I don't see any problem with this pariticular PR, however, deepcopy of a full network isn't a rare usecase which means a fix to deepcopy is needed anyway.

In fact, the behavior of __deepcopy__ of a Block has to be overrided with the _BlockScope being renewed in the new Block, and the suspicious copy behavior of parameters has to be verified and secured.