prevent cyclic Expects warning

[mazunki]

This builds on top of #2363 due to tests.

As requested in #2361 (comment), it's better to handle cyclic faults at the exception handler. This PR fixes this, with some practical compromises.

Ideally, I would have liked to call os::panic(), but this has some issues: currently, os::panic only supports being called with const char *why, with no concept of source location or expressions.

Furthermore, it (used to) assume libc was initialized (i.e. calling printf and friends). I've added a check to prevent this being an issue, irrespective of the problem this PR solves. I think os::panic should have two interfaces: a dumb one that can be used irrespective of initialization state (that is, a simple fallback mechanism that is similar to my implementation of __expect_fail), and a complete one that accepts std::format_string and cooperates correctly with the rest of the system.

Unrelated to this, calling __arch_poweroff() does not feel good. I wanted to call os::shutdown(), but this doesn't actually trigger a shutdown unless the proper message has been sent to vmrunner. I've opened #2364 with more details about this.

[mazunki]

rebased to sync with main

[alfreb]

This looks good - but can you help me figure out how to test it? I will, but might not be today. Also, I agree that we should have an early and late panic, where only the late one expects libc to be initialized- although we could just check if it is. I think there's a boolean for that somewhere.

[mazunki]

We do have kernel::libc_initialized(), but I'm not sure that accounts for all early initialization being ready: 93efffa does exactly that.

I tested this by manually lying in src/kernel/multiboot.cpp::_multiboot_memory_end() (i.e. immediately return os::Arch::max_canonical_addr;) . I'm not sure if there's a clean way to test this, as this should really only trigger if something is /really wrong/ with the early boot process.

[mazunki]

re-based and re-tested.