Extend the TestSparkRunner with its own TransformTranslator containing test evaluators

When we have to test low level features in the Spark runner, requiring access to the `RDDs` or `DStreams`, we have to create a "fake" `PTransform` with a custom evaluators in the core Spark runner.

In order to avoid to pollute the core runner, this should be done in `TestSparkRunner`.

Imported from Jira [BEAM-1051](https://issues.apache.org/jira/browse/BEAM-1051). Original Jira may contain additional context.
Reported by: jbonofre.

Start with TestSparkRunner and the core Spark runner's existing custom evaluator handling. Compare how fake PTransforms and test evaluators are registered for low-level RDD and DStream tests. Done means those test evaluators are provided by TestSparkRunner without polluting the core runner.

Written by the indexing model from the issue text.