Please see branch 'performance'.
I have made a number of tweaks which result in 37% faster parsing of the 1kg VCF (see vcf/test/prof.py).
The breaking change is that most of the time is spent precomputing these. @arq5x this was your change, are you bothered if it goes?
I think this makes sense in the interest of a more broadly useful API. I will need to keep the pre-computation in my fork because of my use case (I need these values for every variant, so pre-computing actually saves me time).