CASSANDRA-21209 Rework ZSTD dictionary compression logic to create a trainer per training

[smiklosovic]

Thanks for sending a pull request! Here are some tips if you're new here:

• Ensure you have added or run the appropriate tests for your PR.
• Be sure to keep the PR description updated to reflect all changes.
• Write your PR title to summarize what this PR proposes.
• If possible, provide a concise example to reproduce the issue for a faster review.
• Read our contributor guidelines
• If you're making a documentation change, see our guide to documentation contribution

Commit messages should follow the following format:

<One sentence description, usually Jira title or CHANGES.txt summary>

<Optional lengthier description (context on patch)>

patch by <Authors>; reviewed by <Reviewers> for CASSANDRA-#####

Co-authored-by: Name1 <email1>
Co-authored-by: Name2 <email2>

The Cassandra Jira

[smiklosovic]

I do not think what we did here was too smart (same concept was there before we started to selectAndReference) because training is done asychronously, so this method returns and finally is called before the sampling is actually finished. We should close in callback, as done above, or only in case we catch exception, as done here.

Or no?

trainer.trainDictionaryAsync(force).addCallback

This "addCallback" makes synchronous call from that? I do not think so, it just registers what should be done after it is finished, but it is not a blocking call, I guess.

[smiklosovic]

whole execution chain (from manager.train) does everything to postpone trainer creation until it is absolutely necessary and all is OK, as the instantiation of a trainer might be memory-wise very demanding (when max sample size is not trivial) as it allocates a direct ByteBuffer. We do not want to create a trainer allocating a big buffer just to throw it away if something else goes south.

[yifan-c]

Some minor nits. Looks good overall. Thanks for the simplification!