Support auditing multi-county contests

[nealmcb]

State-wide contests and many federal and district contests include ballots from multiple counties.

We need better data to audit these contests, in which the CVR either guarantees that contest names are unique across counties, or provides some other sort of unique identifier.

Various options exist for the risk-limiting calculations needed for these contests. A Bayesian Method for Auditing Elections by Rivest and Shen can also handle audits in which some counties are using ballot-polling and others are doing ballot-comparison audits. See e.g. https://github.com/ron-rivest/audit-lab. Other frequentist methods by Stark, at least for the case where all the counties use the same sort of auditing, are available, and more efficient ones are in development.

Given rla_export data from the 2017 election, we can evaluate the efficiency of the various approaches on real-world scenarios and decide how to move forward.

[nealmcb]

Many district contests will also only appear on a fraction of the ballots in a given county. On that front, see also #905.

[nealmcb]

See math for more approaches on calculating risk levels given a combination of ballot-polling audits and ballot-comparison audits, at [1803.00698] Next Steps for the Colorado Risk-Limiting Audit (CORLA) Program