Panic while formatting panic message + always_abort leads to stack overflow

Example and initial analysis by @correabuscar:

#![feature(panic_always_abort)]

use std::fmt::{Display, self};

struct MyStruct;

impl Display for MyStruct {
    fn fmt(&self, _: &mut fmt::Formatter<'_>) -> fmt::Result {
        panic!("this is bad: {}", MyStruct);
    }
}

fn main() {
    std::panic::always_abort();
    println!("{}", MyStruct);
}

The reason for this is that if both "panic while processing panic" and "panic after always_abort" apply, then we use the AlwaysAbort code path which does format the panic message.

This could be fixed by moving this check further down below the in_panic_hook check:

https://github.com/rust-lang/rust/blob/9f25a04498fbe30dcf6a9c764953704c333a0137/library/std/src/panicking.rs#L393-L395

But I don't know if this has other adverse side-effects. Can we even access thread-local variables after fork (i.e., when always-abort is set)?
Cc @Amanieu

Start with the Rust example in the issue and inspect library/std/src/panicking.rs around lines 393-395, especially the AlwaysAbort and in_panic_hook checks. Reproduce the stack overflow, then determine whether moving the check avoids formatting during panic processing without adverse effects, including after fork and when accessing thread-local state.