rolling_mean returns slightly-negative value on Series with only non-negative values

[craustin]

Seeing this in a few cases. This is the simplest repro I can find:

from pandas import date_range, rolling_mean, Series
s = Series(index=date_range('1999-02-03','1999-04-05'))
s['1999-02-03'] = 0.00012456
s['1999-02-04'] = 0.0003
s['1999-04-05'] = 0
rolling_mean(s, 1)['1999-04-05']

Returns 0 in 0.9.0.
Returns -5E-20 in 0.10.0.

We found this issue because we do a sqrt on the result and expected 0s when the input was a 0.

[wesm]

This is related to #2114. I'll have a look

[wesm]

This general problem really sucks. Need a better overall solution to fp error robustness in moving window functions