ENH: L1 penalized trend filtering

[josef-pkt]

this is mainly to park some references

trend filtering is similar in behavior to adaptive splines, but similarly to truncated power splines can be used for sparse penalization, i.e. we can penalize parts/coefficients to zero.

From the approach it sounds similar to the differencing estimators in issue #3384 and PR #3380, but no cross citation in the references.

Tibshirani, Ryan J. "Adaptive piecewise polynomial estimation via trend filtering." The Annals of Statistics 42, no. 1 (2014): 285-323.
https://arxiv.org/pdf/1304.2986.pdf

Wang, Yu-Xiang, Alex Smola, and Ryan Tibshirani. "The falling factorial basis and its statistical applications." In International Conference on Machine Learning, pp. 730-738. 2014.
http://proceedings.mlr.press/v32/wange14.pdf

Kim, Seung-Jean, Kwangmoo Koh, Stephen Boyd, and Dimitry Gorinevsky. "L1 Trend Filtering." SIAM review 51, no. 2 (2009): 339-360.
This article has some implementation in various programming languages on github but essentially all are GPL.
https://github.com/bugra/l1 is Apache, but doesn't look like a lot of code.

(I did not read any of the articles, but saw it briefly described in some course lecture notes by Tibshirani and coauthors.)