Improper log-sum of exponentials in log likelihood

[pblischak]

The log likelihoods of the different models are computed using log sums of exponentials, which can have nasty overflow/underflow problems (cf. this link). I've run into this problem when read counts are high (on the order of ~1000x). Since sequencing coverage isn't typically this high, it hopefully should be an issue for most data sets. Nevertheless, I'll need to fix it.