Do print or use a convergence rate estimate for the first iteration.

The estimate we use can only work starting from the second one.
I hope this patch definitely fixes the problems with spurious NaN, inf,
and astronomical maxTotalConvRates.

[[Imported from SVN: r3040]]