calm() features

[config-i1]

Things to improve in calm():

1. bruteforce=FALSE should get the stepwise model and then do manipulations:
   • the best model + each of excluded variables
   • the model based on purely excluded variables (randomly selected, several models)
   • other permutations?
2. vcov should not be called inside calm, should be done outside
3. vcov.calm should get vcov from each model, not extract the estimated one
4. drop models that have ICw<1e-5 to reduce the pool.
5. pooling maybe?