Thow DomainError when sampling weights contain Inf or NaN

[xiruizhao]

Fixes #671

[nalimilan]

Thanks. Can you add a test?

[xiruizhao]

I am not sure about this now. Should the check be done in weights.jl? Should I also check for negative weights? Or should the library simply ask the callers to ensure valid inputs?

[nalimilan]

Good question. In general it's true that non-finite or negative weights are probably not useful. Though it could be the case that in some applications it's OK to allow them, e.g. if observations with NaN weights should be skipped. That's hard to tell. Anyway that would be breaking so we would have first to print a deprecation warning. In the meantime I think it would be good to merge this PR as it fixes a real bug. But feel free to make a more general PR too.

(BTW, I forgot that the same check should be applied to sample! too.)

[bkamins]

Anyway that would be breaking so we would have first to print a deprecation warning.

I would say it is non-breaking if we decide that allowing Inf or NaN is a bug. I would say it is a bug as I cannot imagine any application when allowing it makes sense. The reason is that we guarantee to calculate sum. And sum is Inf or NaN in these cases - which is not well defined. So, by contract of AbstractWeights allowing Inf or NaN should not be allowed I think.