VeryChord: treat increments below kappa as converged

[raashish1601]

Fixes #782

When the prediction of an implicit stage is already exact, every VeryChord increment is a rounding error. Scaled by tight tolerances these are still well above _small's threshold, so terminate compares the ratio of two rounding errors, often finds it above 2 and reports a divergence. The step is rejected and the same thing happens on the next one, until max_steps.

As suggested in the issue, terminate now also treats an increment smaller than kappa (in units of the tolerances) as converged, whatever its ratio to the previous one. _small is kept for the case where kappa is set very small.

With the reproducer from the issue, Kvaerno5 now takes 52 accepted and 0 rejected steps instead of hitting max_steps (19 accepted, 1981 rejected). The y' = 1.25 y case mentioned there is unchanged (270 steps).

Tests: added test_exact_prediction to test/test_very_chord.py, which fails on main and passes with this change. test/test_very_chord.py, test/test_solver.py, test/test_integrate.py and test/test_adaptive_stepsize_controller.py pass locally (CPU, float64).