Add DREAM synth1 benchmark and variants from VCF

[arahuja]

This PR has a few pieces:

• a benchmark based on DREAM synth1
• loading 'known' variants from a VCF file
• a gcloud configuration (any ideas on how to deal with a reference path here? https://github.com/hammerlab/guacamole/issues/512 could be a fix)
• some py3 fixes (iteritems and 'w' in tmpfiles)

[arahuja]

Not sure if the results make sense for this, I didn't set any parameters, but the summary files gives:

stat,comparison_dataset,numerator,denominator,percent
calls,,1377,,
calls before filtering,,302336,,
calls,published,3537,,
recall with filters,published,50,3537.0,1.413627367825841
precision with filters,published,50,3537.0,1.413627367825841
recall from pooled calling only without filters,published,3508,3537.0,99.18009612666101
recall individual calling only without filters,published,3508,3537.0,99.18009612666101
precision without filters,published,3508,302336.0,1.1602984758679085
precision both pooled and individual triggers firing,published,3508,302335.0,1.1603023136586899

[timodonnell]

Thanks @arahuja . I made a few comments but otherwise this LGTM. From your summary it does look like we way over calling here, will be good to dig into what's happening, may be good motivation for writing a few more filters in joint caller