Atropos

* Abstracted the ErrorEstimator class to enable alternate implementations.
* Added a new ShadowRegressionErrorEstimator that uses the ShadowRegression R package (Wang et al.) to more accurately estimate sequencing error rate. This requires that R and the [ShadowRegression package](http://bcb.dfci.harvard.edu/~vwang/shadowRegression.html) and its dependencies be installed -- MASS and ReadCount, which in turn depend on a bunch of Bioconductor packages. At some point, this dependency will go away when I reimplement the method in pure python.
* The error command now reports the longest matching read fragment, which is usually a closer match for the actual adapter sequence than the longest matching k-mer.

* Bugfixes

* Changed the order of trimming operations - OverwriteReadModifier is now after read and quality trimming.
* Refactored the main Atropos interface to improve testability.
* Added more unit tests.

* Fixed a major bug in OverwriteReadModifier, and in the unit tests for paired-end trimmers.

* Added OverwriteReadModifier, a paired-end modifier that overwrites one read end with the other if the mean quality over the first N bases (where N is user-specified) of one is below a threshold value and the mean quality of the other is above a second threshold. This dramatically improves the number of high-quality read mappings in data sets where there are systematic problems with one read-end.

* Perform error correction when insert match fails but adapter matches are complementary
* Improvements to handling of cached adapter lists
* Merged reads are no longer written to --too-short-output by default
* Many bugfixes and improvements in deployment (including a Makefile)
* Many