add integer programming on x - NB not easy due to tau scaling
cleanup, refactor and comment code
add static, inline to methods
work on conda recipe
multiple passes normalization
adaptive b,c rescaling
fix D, E to be inverse
re-add approximate convergence check
add WARM_START iteration bump so approximate convergence works better in those cases
use python capsules to store matrix fact across multiple runs: https://docs.python.org/2/c-api/capsule.html#capsules

NOTE: higher d->SCALE makes pri resid go down faster, lower makes dual resid go down faster
