Add UI support for min/max aggregation across runs in a trial

The hparams dashboard groups multiple runs with the same hyperparameters
(but different random seed) into a single “trial”. Currently, the
metrics shown in the frontend are always averaged across runs in a
trial. This is perhaps a reasonable default, but it’s somewhat
misleading in that shows metrics that may not be consistent with any
feasible run.

The hparams backend already supports smarter aggregation modes that
select a “best” run from a trial and show consistent metrics from that
run. The best run is chosen by a user-defined metric and aggregation
direction: e.g., ^highest accuracy” or “lowest xent”.

The only problem is that the UI has no controls to actually use this
functionality, so we should add that. :-)

See: API proto definition; backend implementation;
discussion on #1998.

Start with the aggregation modes in tensorboard/plugins/hparams/api.proto and the backend implementation in tensorboard/plugins/hparams/list_session_groups.py, then trace the hparams frontend controls that consume them. Done means users can choose a trial aggregation metric and direction in the UI and see the selected consistent run metrics.