Improve Nuth and Kääb 2011 implementation by stopping when error stops improving

[erikmannerfelt]

Currently, there's a pretty lazy threshold system where if the NMAD drops below 0.05, it will end prematurely:

xdem/xdem/coreg.py

Line 1127 in b22b4b5

if i > 5 and nmad < self.error_threshold:

It would be much better to stop when the NMAD stops improving instead, because most of the time, the NMAD will never reach 0.05. This could be done either with a simple check "check that the NMAD was better than last time" or better, check with the last X errors "check that the NMAD improved compared to any of the last X iterations". The latter is preferable as it is less likely to stop in local minima.

[adehecq]

Solved by #135