Fix loglikelihood calculation with weights

• Fixes #89
• Adds more benchmarks
• Introduces and uses model_response(mf::ModelFrame, d::Distribution) to convert a binary PooledDataArray response to a 0/1 floating point vector. This should be done in DataFrames and a pull request will be made for this.
• pass the random number generator to the simulate! method
• methods for A_rdiv_Bc!, etc. that are now in Base are commented with if VERSION < ...
• allow a contrasts specification in lmm and glmm