Decide inline-worthiness based on a more nuanced cost model

This switches to a model in which we estimate the runtime of the
function; if it's fast, then we should inline it. The estimate of
runtime is extremely crude, and doesn't even take loops into account.