Documentation: guidelines for interpreting R^2 and stddev

In running the ~2000 criterion benchmarks currently on stackage, we've run into the problem that some of them are _clearly_ nonlinear.

We are trying to develop a data cleaning step which will throw these out before comparing two different runs of the full benchmark set. Once we find good thresholds, I would like to put some guidelines together to share these with criterion users.

This goes along with other documentation updates, e.g. #92 and #95.

CC @RyanGlScott @vollmerm