Add support for parallel composition in dp-accounting

In some mechanisms it is helpful to partition the dataset and run a different mechanism on each subset. The privacy properties are then the worst-case across all subsets. It would be good to have support for this in dp-accounting.

Start by locating the dp-accounting package and reviewing its existing composition support. Clarify how partitioned datasets and different mechanisms should be represented, and how the worst-case privacy result should be exposed. Done means the intended parallel-composition behavior is specified, implemented, and verified for relevant mechanisms.

Written by the indexing model from the issue text.