fix rare duplicated data buffer entry

[youliangtan]

Issue

In bad network connection, agentlace's TrainerClient might not track the latest inserted batch data correctly, which resulted in duplicated data entry from the actor node to learner node. This will degrade the performance of online learning, since the sequential nature of the data is no longer obeyed.

Fix

[Highly Recommended]

The straightforward fix is to bump up the version of agentlace (or pull the latest main). For more details, checkout the PR: youliangtan/agentlace#11

Additional

1. The current update() method uses a pure service request response method, it is blocking and might not be fast enough. An experimental mode with experimental_pipeline_url experimental_pipeline_port method is introduced to speed up the method.

Related to youliangtan/agentlace#13

2. Reduce default size of QueuedDataStore to reduce unnecessary memory consumption on the actor node
3. Misc clean up of code

[youliangtan]

@charlesxu0124 @jianlanluo FYI