Allow y to be a big.matrix instead of a r vector

[coatless]

The restriction that y is an r vector hurts the big application of this package. If I'm working with over 45 million observations, then the y vector of only 0s and 1s take up about 180 mb vs. practically nothing under bigmemory object.