Streaming Beam SQL benchmarks on all of our runners

Beam has a number of classic streaming SQL benchmarks known as "Nexmark" coded up in both raw Java and also Beam SQL.

So far, expanding functionality has been the focus of Beam SQL so there is little known about performance - we know only that it is a pretty straightforward mapping from SQL to Beam that should work OK a lot of the time. It would be interesting to see where the bottlenecks are when these SQL benchmarks are translated via Beam SQL into a Beam pipeline and then again translated to the native capabilities of e.g. Spark and Flink.

This project will require the ability to read, write, and run Java and SQL.

Imported from Jira [BEAM-3783](https://issues.apache.org/jira/browse/BEAM-3783). Original Jira may contain additional context.
Reported by: kenn.

Start by locating the existing Nexmark implementations in raw Java and Beam SQL, then identify how benchmarks are launched across the available runners. Compare the Beam SQL pipelines with native Spark and Flink execution, and consider the work done when benchmark results and bottlenecks are reported for all runners.

Written by the indexing model from the issue text.