BUG: stats: don't emit invalid warnings if moments are infinite

[pv]

Silence nan floating point warnings in rv_generic.stats, when
calculating derived values that may be nan if higher moments are
infinite.

This means calculations that do inf - inf or inf/inf.

Fixes the one crystalball moments failure currently on master.

[josef-pkt]

Do you have a link to the test run to see which distributions this affects?

Distributions that have undefined (nan) or infinite theoretical moments should specify this in the distribution specific method.

If these are numerical problems like overflow, then maybe the examples should be adjusted.

(I don't remember anything about crystalball)

[pv]

https://travis-ci.org/scipy/scipy/jobs/259105158

[pv]

The distribution I think explicitly returns infinities from its _munp method as appropriate.
But stats() does further computations based on these values.

[josef-pkt]

Ok, it would be possible to avoid computing g1 in the generic code if either m2 or m3 or both are inf or nan.
However that requires another lazywhere or similar while floating point computation with it produce the correct results with just the warning noise.

[pv]

Yes, as far as I understand the produced results are reasonable, and it would be a bit fiddly to explicitly specify these when whether the moments exist depends on the parameters.

[ev-br]

One could argue whether this should be fixed at the level of the concrete distribution (cf_lazywhere gymnastics at crystalball._munp). However this is a simpler fix, and in it goes.