Add support for InMemoryRelation

[sunchao]

What is the problem the feature request solves?

Currently Comet cannot be triggered if Spark users read data from cached RDD. To support this use case, we'll need to add support for Spark's InMemoryRelation.

It looks like we may need to implement Arrow for CachedBatchSerializer.

Describe the potential solution

Add Comet support for InMemoryRelation, so that Spark query starts from cached RDD can also use Comet native execution.

Additional context

It is not a priority as of now, but will be something good to have in future.

[advancedxy]

Another way to read InMemoryRelation is to wrapped it with an CometRowToColumnarExec like I proposed in #119

[sunchao]

@advancedxy Yea, CometRowToColumnarExec could be a more general solution, not only for InMemoryRelation, but also for other types of data sources like CSV, JSON, etc. The advantage of implementing Arrow for CachedBatchSerializer here is that we can avoid the extra cost from row to columnar conversion, and potentially be more space efficient because of better compression.

[advancedxy]

The advantage of implementing Arrow for CachedBatchSerializer here is that we can avoid the extra cost from row to columnar conversion, and potentially be more space efficient because of better compression.

Yea, of course. I can get the rational. We could always add specialized operators to improve performance as long as it's worth the effort and there's interest to implement it.