Scale var/std

[PokhodenkoSA]

Only var is implemented yet.
This PR is based on #610 because used parallel nanmean.

std autoscaleup because Series.std is implemented via Series.var. @densmirn thank you :)

[AlexanderKalistratov]

You could optimize it further. The formula for variance is:
(1/n)∑(aᵢ - mean)² = (1/n)∑(aᵢ² - 2aᵢ\*mean + mean²) = (1/n)(∑aᵢ² - 2\*mean\*∑aᵢ + mean²*∑1)
mean is:
mean = (1/n)*∑aᵢ
Which give us the following:
(1/n)(∑aᵢ² - (2/n)∑aᵢ*∑aᵢ + (1/n)²*∑aᵢ*∑aᵢ*(n)) = (1/n)(∑aᵢ² - (2/n)*∑aᵢ*∑aᵢ - (1/n)*∑aᵢ*∑aᵢ) = (1/n)(∑aᵢ² - (1/n)*(∑aᵢ)²)

So, you could implement variance as:

square_sum = 0.
sum = 0.
total_count = 0
for i in prange(len):
    a = self._data[i]
    if not isnan(a):
        square_sum += a*a
        sum += a
        total_count += 1

if total_count < 1:
    return numpy.nan

return (square_sum - sum*sum/total_count)/total_count

Also you could see covariance as an example

Also, haven't validate the final formula. So there could be errors

[PokhodenkoSA]

Also, haven't validate the final formula. So there could be errors

I will implement it in separate PR.