Add dedup flag to master from #19112

[samskalicky]

Description

Added option to disable deduping the subgraph inputs/outputs. @bgawrych asked for this while they migrate the MKLDNN quantization flow from using executor to optimize_for.

Original PR for subgraph deduplication on master #16131. When backporting to 1.x we added the option to enable deduplicating subgraph inputs/outputs in #19112. On master the option will be enabled by default, and the flag will allow disabling as needed.

[mxnet-bot]

Hey @samskalicky , 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: [edge, clang, website, centos-cpu, sanity, windows-gpu, centos-gpu, windows-cpu, unix-gpu, miscellaneous, unix-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.

[mseth10]

LGTM

[bgawrych]

LGTM, thanks
@anko-intel @pengzhao-intel Maybe you want to take a look too

[grygielski]

@samskalicky I've got a question about the default value of dedup_subgraph flag. It is set inside the constructor to true. However if user does not provide {'dedup_subgraph': 'True'} to option_map, it is set to false anyway. Is it desired outcome? It looks like a bug to me.
https://github.com/apache/incubator-mxnet/blob/0514233103baff5e1581cf2057f561f7a36616c2/src/operator/subgraph/subgraph_property.h#L271-L279

[samskalicky]

@samskalicky I've got a question about the default value of dedup_subgraph flag. It is set inside the constructor to true. However if user does not provide {'dedup_subgraph': 'True'} to option_map, it is set to false anyway. Is it desired outcome? It looks like a bug to me.
https://github.com/apache/incubator-mxnet/blob/0514233103baff5e1581cf2057f561f7a36616c2/src/operator/subgraph/subgraph_property.h#L271-L279

Thanks @grygielski for pointing this out. This does seem to be the default case. However, for the subgraph property that is used in the optimize_for flow it does not implement this logic in its PrePartition function:
https://github.com/apache/incubator-mxnet/blob/0faecf01eddc781793927cb7cb9415310e9b21a2/src/operator/subgraph/partitioner/custom_subgraph_property.h#L191-L195
So while the default case for the parent function of the SubgraphProperty class does this logic, subclasses that implement their own PrePartition function can override this behavior as needed.

I should probably clarify that "being enabled by default" applies to the optimize_for flow, rather than all subgraph properties being enabled by default. Sorry for the confusion.