Allow switching between accurate vs probabilistic model in pipeline

[chubei]

We're using probabilistic models in the pipeline to improve performance. There are two patterns:

• Using hash of something as the key of a HashMap, it appears in aggregation processor and join processor.
• Using a bloom filter to test for existence. It appears in set processor.

We'd like to enable the users to choose between accuracy and performance. This can be a runtime flag the the whole pipeline level. The individual processors will respect this flag and use corresponding implementation.

This requires:

• Implement "accurate" model for relevant processors.
• Add the flag to the configuration file.

Implementation note: We should probably clean up join processor before implementing the change. There are quite some duplicated code.