0.12.0

Added

• First part of decay in fast binless is not forced to decrease. Use free_diag
  argument to decide (default 10kb)
• patchno column to fast binless output
• allow to pass a vector of lambda2 values to fast binless
• allow to pass lambda1 to get thresholdings similar to optimized binless
• allow to use final output of optimized binless as input to fast binless

Changed

• Exposures are now fit separately
• Use nobs column in fast binless
• ncounts becomes nobs in optimized binless
• Use same column names in fast and optimized binless outputs
• fast binless model now mimicks that of slow binless
  • using binwise averages (breaks backwards compatibility)
  • using negative binomial with fixed dispersion
• require pos1 and pos2 columns in fast binless input
• fits in optimized binless are performed by group and are more efficient
  memorywise
• drop stan dependency and fit dispersion manually on a subset of rows, taking
  the median value
• Switch to BIC to estimate lambda2
• reorder arguments in fast_binless

Fixed

• sigma parameter was ignored in GeneralizedAdditiveModel
• output matrix properly returns factor labels for name, bin1 and bin2 in fast binless
• bug causing weights to be twice too small
• centering bug on first iteration of fast binless
• bugs in the BIC calculation of optimized binless, causing lambda2 estimates to
  be too low and unstable
• Bug in fast binless when entire rows had nobs == 0