add end2end testing

[martinjaggi]

document for us internally the key metrics for some of the standard tasks (e.g. titanic), and test those by hand before/after changes which could affect it. can be done by human, does not need to be automated. this is important in addition to the unit tests.

1. test accuracy for training alone
2. test accuracy for 2 clients, with the same dataset strictly partitioned between the two

no code needed for now but clear values needed and file splits and documentation of expected results in a readme etc (or on the task)

ideally we'd also have these 2 models tested on a completely separate holdout set, removed from the training set before given to disco.

will make a new issue to support validation on separate set