Fix reductions in Statistics

[timholy]

Formerly we were choosing an inconsistent eltype for reductions performed by Statistics. Since that's a stdlib that comes for "free," it seems we should support it on similar footing to reduce etc.

[codecov]

Codecov Report

Merging #183 into master will decrease coverage by 0.20%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master     #183      +/-   ##
==========================================
- Coverage   87.84%   87.64%   -0.21%     
==========================================
  Files           5        5              
  Lines         436      437       +1     
==========================================
  Hits          383      383              
- Misses         53       54       +1     

Impacted Files	Coverage Δ
src/FixedPointNumbers.jl	80.64% <0.00%> (-0.88%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0872417...c9ca80b. Read the comment docs.

[timholy]

Any objections or should I merge?

[kimikage]

@timholy, are these #183, #180 and #177 in v0.8.1 v0.8.2? Or in the v0.9 series?

As you mentioned, #177 is a somewhat dangerous change, but it shouldn't affect the "normal" functionality of downstream packages if there is no type piracy.

#186 is not a breaking change, and is not urgent. #187 is somewhat dangerous because it prohibits copysign/flipsign for Normed.

[timholy]

#180 is in 0.8.1. None of the rest have been released yet. I think everything merged can be in the 0.8.x series, though, see #177 (comment).

[kimikage]

Ah, I've already used v0.8.1, but I forgot it. 😅
My interest is when to release v0.9. Originally in v0.9, I was planning to improve arithmetic.

[timholy]

I'd say we could make a 0.8.2 release and then add any breaking changes.

Where do we stand on overflow? Without a clear consensus I'm reluctant to pick #143 up again. There's a part of me that thinks we should make all arithmetic immediately promote to floattype and just declare that these are storage-only numbers.

[kimikage]

I will not drastically change the promotion rules unless there are concrete proposals. In other words, there has been no debate in the last six months, so I want to improve the arithmetic.

My first goal is to complete the set of checked_*, wrapping_*, and saturating_*.(cf. #152 (comment))

Edit:
I submitted a draft PR #190.