Michael, here are the changes to make statistics optional.

There are two issues I want to note.

First, you and I talked about statistics_log implying statistics, and I ended up going the other way. The reason is because having the statistics boolean affect statistics_log means you can configure statistics logging, and then turn it on/off using the statistics boolean. That seemed more intuitive and simple to me; if you disagree, let's talk it over and I'll fix it.

Second, there's a nasty side-effect of this change. We now have two classes of statistics, those available when the run-time statistics aren't configured (btree file row count, for example), and those that aren't. I didn't try to address the problem in this change, I wasn't sure how to solve it: we could ignore the problem, we could document it ("some statistics are always available", or "some statistics are run-time only", here's a list), or we could split up the structures into run-time and not-run-time, and configure the stat cursor to return one set or the other (we already have something a little like this in "statistics_fast").