Complete rewrite of bottleneck package

[kwgoodman]

I have completely rewritten the bottleneck package. Since it is an optional dependency of pandas I thought you might want to test it before I release.

Release notes are here: https://github.com/kwgoodman/bottleneck/blob/master/RELEASE.rst
Branch: master

Please report results here: pydata/bottleneck#96

[shoyer]

The only breaking change I noticed for pandas is that the new version of bottleneck changes the behavior of nansum for all NaN arrays to return 0, rather than NaN. Otherwise the test suite ran successfully (see the linked bottleneck issue).

This shouldn't be a blocker for the bottleneck release but we should address this nansum change for pandas 0.16 (see #9422).

[jreback]

1.0 pretty popular now. so all except for the nansum change #9422 this issue not relevant anymore.