[FLINK-27286] Add infra to support training high dimension models

[zhipeng93]

What is the purpose of the change

This PR introduces a communication infra that can distribute model parameters to multiple servers, so as to support training high dimension models in Flink ML. Moreover, it also introduces a programming API that can simplify the programming of iterative machine learning training process. The motivation cases include:

• Train a high dimension model whose model data cannot be fitted on a single machine.
• Support the case where the model data of each key as a double or double[].
• Support model data with discrete index.
• Use a set of index to retrieve the model values.
• Use a set of index and UDF to retrieve the statistics computed from model values.
• Use a set of index/value to update the model values.
• Use a set of index/value to update the model, with user specified aggregation logics.
• Supports AllReduce/ReduceScatter for java generics.
• Supports AllReduce/ReduceScatter to execute every K iterations.
• Support the case that user can access the training data in each iteration.
• Output model data to downstream tasks.
• Output the intermediate result to downstream tasks.

Brief change log

• Introduced WorkerOperator and ServerOperator to execute the training logic and maintain the model data, respectively.
• Added MLSession to store the information that is alive during the training process on WorkerOperator.
• Added Message to represent the information that can be transfered among workers and servers.
• Added IterationStageList and IterationStage to describe the iterative machine learning process as a chain of computation/communication stages. Also, computation stage (ProcessStage) and commonly used communication stage (Push/Pull/AllReduce/ReduceScatter) are also added.
• Added ModelUpdater to let developers describe model updating logic.
• Added TrainingUtils to ease the programming of iterative machine learning process.
• Added SharedLongArray and SharedDoubleArray to allow developers reuse the memory accross different iterations.
• Added unit test to cover the above functionalities.

Does this pull request potentially affect one of the following parts:

• Dependencies (does it add or upgrade a dependency): (yes)
• The public API, i.e., is any changed class annotated with @Public(Evolving): (yes / no)

Documentation

• Does this pull request introduce a new feature? (yes)
• If yes, how is the feature documented? (JavaDocs)

[zhipeng93]

Hi @lindong28 @Fanoid @weibozhao @jiangxin369 , can you help to review this PR?

The PR before [1] is a bit complicated since it involves too many changes (i.e., infra, vectors, algorithms). So I extract the infra here as a separate PR. The comments left in [1] are addressed in this PR.

[1] #237