Speed up GLMM mode

[jr-leary7]

• remove fitting of the subject-specific GLM models within fitGLMM()
  • this requires som re-working of various bits and ends but should be feasible
• see if it's possible to speed up the NB LASSO bit, as per profiling with profvis that is the largest runtime piece
  • maybe supply a smaller / more intelligent set of values of the penalty parameter $\lambda$ ? right now 50 values are considered, but the values themselves are chosen automatically

[jr-leary7]

Also figure out whether MNET penalty actually induces sparsity or not ...

[jr-leary7]

Dumb question -- MNET does induce sparsity of the coefficients a la LASSO. maybe add an argument to fitGLMM() specifying whether LASSO, MNET, or SNET penalties should be used ? do an analysis to determine which results in "most sparse" outcomes, as that's likely what we want