Fix 'auto' encoded longs + compression serializer

[clintropolis]

Fixes #6044 by changing BlockLayoutColumnarLongsSerializer block flushing to call LongEncodingWriter.setBuffer which will create new serializers.

To help catch further issues like this, the only thing I could think of that would make me feel reasonably confident about stuff was to add a rather exhaustive test that tries a variety of 'bits per value', number of rows, and compression strategies, but it unfortunately adds a minute or two to overall test time.

Running io.druid.segment.data.CompressedLongsAutoEncodingSerdeTest
Tests run: 104, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 58.116 sec - in io.druid.segment.data.CompressedLongsAutoEncodingSerdeTest

It isn't even completely exhaustive, as the behavior of IntermediateColumnarLongsSerializer isn't directly controllable, but both table and delta encodings are exercised. I'm considering marking it @Ignore.

[gianm]

I don't think the test you wrote should be @Ignore -- it is valuable to have some exhaustive tests that we run before releases. Does JUnit have a way to tag tests as "slow" and then only run them either for releases, or when you specifically ask for them?

[jihoonson]

I think it's fine to enable by default if it just takes one or two minutes since this kind of bugs is difficult to catch without various testing.

[jihoonson]

@clintropolis what do you think about adding flush() to LongSerializer?

[clintropolis]

@jihoonson I considered that but it seemed more invasive of a change and I'm trying to deprecate this code path anyway..

That said, it is strange that close is called multiple times, so I'll just go ahead and refactor.

[clintropolis]

@jihoonson What if I just rename LongSerializer.close to LongSerializer.flush and if it is tied to an output stream it can only be flushed once, leaving the code structurally as is?

I think alternative involves adding a close method to LongEncodingWriter and a flush method to LongSerializer and changing the hooks into LongEncodingWriter from EntireLayoutColumnarLongs to call close instead of flush

[jihoonson]

@clintropolis I'm fine with it.

[jihoonson]

The latest change looks good to me.