Extend test coverage with noisy time series

[brandtg]

We should add a test that generates some series with known trend + seasonal and noise / anomalies, then see if we can decompose and get something similar w/in error bounds.

Something similar to this but maybe less simplistic: https://github.com/brandtg/stl-java/blob/master/src/test/java/com/github/brandtg/TestStlDecomposition.java#L42

Alternatively, we could find some good public time series data sets with noise (e.g. the CO2 level data from R that is in sample-timeseries.json).