Specifying a negative tolerance now lets you avoid trying to directly
calculate or guess what a decent L2/H1/whatever error growth rate is,
by using a multiple of the growth rate observed at the initial time
I haven't yet tested this, just thought I'd let folks take a look at the idea while I do.
Allow TimeSolver tolerances relative to first step
Better AdaptiveTimeSolver documentation
Paul pointed out the need for documentation. If nobody else has concerns I'll merge this tomorrow.