Reductions shouldn't overflow for LinearSlow arrays

[timholy]

Talk about embarrassing:

julia> A = reshape(map(UInt8, 101:109), (3,3))
3x3 Array{UInt8,2}:                                                                                                                                                                                                                                                            
 0x65  0x68  0x6b                                                                                                                                                                                                                                                              
 0x66  0x69  0x6c                                                                                                                                                                                                                                                              
 0x67  0x6a  0x6d                                                                                                                                                                                                                                                              

julia> sum(A)
0x00000000000003b1                                                                                                                                                                                                                                                             

julia> sum(sub(A, 1:3, 1:3))
0xb1

I can't believe no one (first and foremost, myself) has caught this; we've had it since #8420 was merged. As someone who has been writing/prodding at related code for quite a while, I am ashamed.

[timholy]

@tkelman, we should let this run through PkgEvaluator, but assuming all is well this is a pretty high priority backport (especially with the tagging of the new FixedPointNumbers, which is how it got revealed). Let me know if there's anything you want me to do to help.

[tkelman]

I'll be starting work on a backport branch within the next week, so thanks for the heads-up. What will certainly save me the most time is a few other people trying to run PackageEvaluator locally and seeing if anyone other than me hits the issue described in #13718 (comment) - getting to the bottom of and fixing that will reduce the number of times I have to repeatedly run PkgEval.