Use more accurate statistics for cursor cache totals

The current statistics used are not modified atomically, and are probably not so useful to use for a resource counter.  Over time, cache effects may cause a drift in the value.

At the minimum, the current global count of cached statistics should be accurate (that is, incremented/decremented in an atomic way.  If there is a measurable performance hit by adding this, we could avoid changing the open_cursor_count metric when caching/reopening a cursor, and change the meaning of that metric to include cached cursors.