Implement ballot-comparison audits

[nealmcb]

The paper A Gentle Introduction to Risk-limiting Audits introduces ballot-level comparison RLAs.

Re: #88: Implementing BRAVO, for ballot polling audits, will be appropriate at some point. But ballot polling audits are far less efficient. When we have CVRs, what we want is ballot-comparison audits, as referenced and implemented for example at https://github.com/nealmcb/audit_cvrs/blob/master/audit_cvrs/rlacalc.py and in Stark's code via javascript in this repo at StarkRLAJS/javascript.js.

Use the "nmin" function (of either codebase) to calculate whether we've hit the risk limit,
given observed discrepancies, and the nminFromRates function to estimate an initial sample size
from discrepancy rates.