Add first value and last value aggregations

_From @shaycrk on August 31, 2017 22:8_

In some cases, we may want to build features that reflect the earliest known or most recent information about a characteristic such as charges at first arrest, current employment status, outcome of last inspection, etc.

The built-in postgres aggregate function (e.g., min, max, avg, etc) don't make this possible but `first` and `last` aggregates are available through a [postgres extension](https://wiki.postgresql.org/wiki/First/last_(aggregate)), which can be invoked with an `ORDER BY` clause in the aggregate call (e.g., `LAST(employment_status ORDER BY knowledge_date ASC)`.

I don't think it would be difficult to implement this here and would be useful, but how do we feel about providing options that rely on postgres extensions rather than built-in functionality?

_Copied from original issue: dssg/collate#97_

No implementation files, tests, or entry points are named. Start by locating the existing aggregation implementation and reviewing the linked PostgreSQL first/last aggregate extension, then clarify the supported API, extension dependency, ordering behavior, and test expectations. Done means the project has an agreed implementation scope for first-value and last-value aggregations.

Written by the indexing model from the issue text.