[Fix] Several bugs when using HvdAllToAllEmbedding to train model and save/restore KV files. #1629

Job	Run time
Bazel code format	8s
Check that we can build the docs	9m 12s
Notebook format	17s
Clang C++ code format	13s
Notebook lint	16s
Check that the pre-commit works	41s
Check that the source code is compliant	9m 55s
Run the cpu tests in a small python docker image	22m 24s
Valid build files	1m 35s
Check that the CODEOWNERS is valid	10s
Yapf Python code format	46s
	45m 37s

Provide feedback