Fix fallback for errors in logging

[danielwe]

This fixes the bug identified in #3012, in which the catch block that's supposed to save you in case an error is thrown during logging ends up throwing an error of its own.

• original_stderr is now assigned at runtime through the module's __init__ function, not at compile time as a hardcoded global constant. This was the source of the bug.
• I rewrote the catch block's print/showerror as an @error log, which stands out more with the red frame and also looks more consistent with the other information logged to the REPL by Pluto.
• Capturing stderr output for testing is not entirely straightforward. There are tradeoffs between using a local and remote worker. Please review the comments in the testing code and let me know which approach is preferred (or if there's a better design I haven't thought of).

Having looked more deeply at #3012, I think this PR is a sufficient fix on the Pluto side. The reason GPUCompiler's log messages error in the first place is that the purportedly safe logging macros are implemented incorrectly and aren't actually safe. I'll put together a PR to GPUCompiler for that.

[github-actions]

Try this Pull Request!

Open Julia and type:

julia> import Pkg
julia> Pkg.activate(temp=true)
julia> Pkg.add(url="https://github.com/danielwe/Pluto.jl", rev="loggingerror")
julia> using Pluto

[danielwe]

Apparently, capturing stderr from the remote worker doesn't work on either Windows or julia 1.6; the latter because redirect_stderr(::Pipe) isn't supported in that version, the former for unknown reasons. I guess a local worker is the way to go for testing then. The latest commit implements that.

[fonsp]

Thanks for the research and the nice solution! Looks good!

(I changed the PR a bit to move the redirect_original_stderr functionality from the PlutoRunner source to the test.)