Fix computation of range length in certain cases (fix #20373)

[StefanKarpinski]

No description provided.

[tkelman]

why the skipped tests? x-ref to the specific comment describing them #20373 (comment)

[StefanKarpinski]

Ok, there's a real 32-bit failure here on Linux & Windows. The stack traces seems to be junk, but the error is an InexactError, which may be from the floor(Int, ...) call. That was simply moved to a later point, so maybe it's something else, but this also adds tests that weren't being run before.

[StefanKarpinski]

32-bit failure is in the linspace constructor for random ranges, cc: @timholy. I'm just going to skip those tests in this PR since I haven't touched linspace at all, but I'm leaving a FIXME note and we should open a corresponding issue to fix this issue since the tests I added reveal a failure case.

[StefanKarpinski]

This PR should be squashed when merged to avoid a broken 32-bit commit in history.

[StefanKarpinski]

AppVeyor failure is unrelated, occurring after normal tests have passed (the ColorTypes thing).

[timholy]

Sorry, I didn't follow this in real time. What was the linspace error? The links here to the test aren't revealing.

[timholy]

Boy do we need #20005. It's one of those things that takes one person much more time than they get back, but collectively...sorry I haven't gotten around to finishing it.

[StefanKarpinski]

I opened an issue here: #20520. The backtraces are mostly unhelpful – they simply indicate some unknown InexactError that occurs in that block of tests. More informative is that fact that the failure goes away when I don't call linspace. Unfortunately, that's all I've got.