Restore save/load ndarray to 1.4.1

[reminisce]

Description

This PR restores the NDArray::Save/Load functions to 1.4.1 for keeping backward compatibility for saving/loading zero-size tensors. The issue was reported in #14954 (comment) by @DickJC123.

This PR depends on #15063. I will rebase this PR after #15063 is merged.

@DickJC123 Since we are not going to have the NumPy shape semantics in the next release, I simply reverted the changes for serialization functions introduced by #14661. After careful thinking, I find that it's required to depend on the compatibility mode to save/load ndarrays in legacy/numpy shape semantics (we will also need a magic version number to guard this when loading). We will deal with it after introducing NumPy operators. Right now, I added CHECK in save/load functions indicating that numpy shape semantics cannot be turned on.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

[marcoabreu]

We have a backwards compatibility test running constantly. Could you check why it hasnt caught that?

[reminisce]

@marcoabreu This is specifically about serializing/deserializing zero-size tensors, which are the kind of tensor not supported by any operators in MXNet. For this reason, there were no such test cases in MXNet before, but sockeye has a unit test covering this case. I added a similar test in MXNet to exclude possible incompatible changes in the future. Thanks.