fixed constants detection. -there was the famous `ramp-up` error, because of the laggy implementation of pd.rolling ;)

advanced test suit... failing tests have reasons eg. skew need 3 min_periods, kurt at least 4, etc. count produce NaN, where pandas.rolling produce 0.0, due to a pandas internal function call to .count if a indexer is given (freaky!)..

absolutly pandas conform tested - forward also works like pandas forward-hack, center is also tested