Compress TDigest can generate centroid with NaN mean

[tisonkun]

See this thread apache/datasketches-rust#23 (comment)

Also we don't verify that the centroid mean is not NaN and the weight is not zero (should be positive even?) on deserialization. (Generally we can't assume the data to be deserialized is safe.)

I think the rank/quantile method can return NaN if the calculation result in that. But the stored centroid should not be NaN or else stable sort is unsound?

[tisonkun]

Even we filter all inf/-inf/nan, it's possible that f64::MIN - 1 produces -inf and later result in NaN.

I noticed the cause is from our calculation in the new mean:

        centroidMeans_[numCentroids_ - 1] += ((values[current] - centroidMeans_[numCentroids_ - 1])
            * weights[current]) / centroidWeights_[numCentroids_ - 1];

And in apache/datasketches-rust@b35cdb2, I change the calculation from:

m1 + ((m2 - m1) * w2 / (w1 + w2))

to

m1 * (w1/(w1 + w2)) + m2 * (w2/(w1 + w2))

... with calling mul_add to avoid double precision as much as possible. Hopefully this would eliminate all the NaN possible sources.

[leerho]

Sounds like this should be a bug fix for both C++ and Java.