Make mcmc_lc accept log priors, not prior pdfs, to avoid underflow

[dannygoldstein]

This PR alters the interface of the mcmc_lc method, making its priors argument take a dictionary of functions that return the log prior, rather than the prior pdf. Reasons for doing this:

• computing the log probability directly is often cheaper than computing the pdf and then taking the log
• it avoids underflow (no situations where an exception will be raised by trying to take the log of 0)

[kbarbary]

This looks like a good idea. But we should make it backwards-compatible. How about adding a separate keyword like lnpriors or logpriors that takes precedence over priors if specified?