Implement rounding in length(::StepRange)

[timholy]

StepRange is the only range object currently in Base that's suitable for non-Integer, non-AbstractFloat objects; among possibly-other places, it's used to construct ranges with physical units in Unitful. I noticed an off-by-one bug in length with such ranges, which this fixes. CC @ajkeller34.

[timholy]

I made a change to range that also makes it more robust against roundoff error. I did this as a deliberately-conservative change, aka this implementation is not very efficient but it's guaranteed not to break unless we were already getting the wrong answer (and many cases where we were getting the wrong answer will be handled correctly).

Looking it over the first commit again I noticed that this change would be breaking for a type that implements div but not /. While I suspect this would be unusual, it seems possible. (I think Dates might be an example, although Dates "rolls its own" when it comes to length which is why this change passes our tests.) Before merging, is it possible to find out whether whether this PR would break anything with a PkgEval run?

[tkelman]

yes. the process for that is.described at #17165 (comment)

[timholy]

Now that we have generic LinSpace and StepRangeLen, I think this isn't so necessary anymore.

[tkelman]

relevant for release-0.5 or not?

is there any overlap with #20396 ?

[timholy]

Seems like it should be OK on 0.5; the only thing I can see to be concerned about is whether the constructor might take longer to run than it used to.

There's no overlap with #20396, because that only affects FloatRange whereas this is about StepRange. In 0.5, StepRange is the only type you can use to create, e.g., Unitful ranges. Those don't get caught by the complain-if-you-throw-me-an-AbstractFloat test, either, yet you can get corruption in some cases.