Safe LayerNorm

[sxjscience]

Description

Change all reduce in the LayerNorm operator to the "safe" version introduced in #14616. In essence, the safe version uses a higher-precision dtype to store the result of reduction. This may avoid the non-convergence problem in training Bert using Float16.

Partially solves #14073.

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

Changes

• Always use the safe reduction in LayerNorm

Comments

• Similar to Fp16 support for layernorm #14073, the change is backward-incompatiable. The results of LayerNorm will be different when the input type is float16.

[sxjscience]

The error in Julia /~https://github.com/apache/incubator-mxnet/blob/master/julia/test/unittest/ndarray.jl#L900 is not related to this PR.

[Roshrini]

@anirudh2290 @apeforest Can you help review this PR?

[eric-haibin-lin]

Are we relying on any exiting unit test?

[roywei]

@sxjscience Could you rebase your PR, the Julia CI failure should be fixed now. Thanks!

[eric-haibin-lin]

We need to make it backward compatible like #14830