ENH: outlier robust seasonal_decompose

[josef-pkt]

related #4044 using non- or semi-parametric methods

(just an idea, no references for now)
Replace the estimators for the components with outlier robust methods.
This is a similar idea as using lowess for the trend, i.e. robust non-parametric smoothing, but also applied to seasonal pattern.

seasonal_decompose uses filters that are not robust to outliers, e.g. mean for seasonal component and convolution filter for trend. Outliers, either isolated or an episode or outliers can or will distort the seasonal pattern and the local trend.

I have seen some plots on stackoverflow or blog post where this seems to happen.

Additionally, the default window length looks often too small, and I haven't seen users overriding the filt window. A short window increases the local effect of an outlier, although it doesn't spread out the effect as much as a larger window. Also short term fluctuations over several periods will be reflected in the local trend.

One possibility to make changing and experimenting with the window length easier, would be to accept an integer for filt that defines the window length used with the current uniform kernel.