Support for attribute slicing in benchmarking module

[andkay]

In previous iterations, we required project-specific implementations of benchmark tools. The more generalised approached is better but is problematic when benchmark data is logically related to a subset of the entire population.

The particular example that is causing a headache at the moment is trip_mode_shares_comparison. Typically, the benchmark data would come from census journey to work data or similar. The comparison requires looking only at modeshares for the "general population" -- i.e. excluding goods vehicles and agents that are external to the study area (which we include sometimes to get the full demand on the local network, but don't actually care about otherwise).

There are some workarounds, including restructuring the benchmark data to be grouped by subpopulations (with arbitrary values) or by using a destination_activity_filter to exclude particular activities (e.g. delivery). But these are cumbersome and potentially limit our ability to score simulations on how close they match benchmarks.