Fix #9209

[ivarne]

This fixes #9209, which is a regression from #9137. The problem was that the Latex completion code
might access an invalid index in the unicode string, if the previous
character from a \\ was a multibyte character.

cc @tkelman @Luthaf @dhoegh

[tkelman]

Did you intend to cancel the appveyor run? As @Keno said here #9209 (comment) the fix for this should come with a regression test.

[ivarne]

Yes, it ran once and failed on 64 bit, and I tried to rerun only that test. AppVeyor decided to run both again, so I canceled, but the link was already updated. I agree about a regression test, so I haven't merged this yet. I'll have to dive somewhat deeper into this to understand how I might write the test.

[dhoegh]

You could make the regresion test like s = "α\\alpha" on https://github.com/JuliaLang/julia/blob/master/test/replcompletions.jl#L70

[tkelman]

@dhoegh useful trick when you're referencing lines of code, hit y when you're viewing the file and github will reload the page with the current sha in the link, that way the same line of code will make sense in the future

[ivarne]

Sorry that it took so long to write the test. Now it's there and passes.

[tkelman]

Was this not fixed by #9227 ?

[stevengj]

@tkelman, no, that was a separate problem.

[ivarne]

Backported in 6033478