bpo-34481: Fix surrogate-handling in strftime

[pganssle]

This is intended to be re-based against #8878 - that PR adds tests, this one adds the fix.

I am making this PR before that one is merged mainly because this is some crazy platform-specific behavior, and I want to test it on the full cross-platform test suite.

https://bugs.python.org/issue34481

https://bugs.python.org/issue34481

[pganssle]

I have no idea how to reproduce this failure. It works fine on my home computer (Arch Linux), on a debian VPS and a debian WSL. All are succeeding.

I think that this will start succeeding on Linux as well if I re-enable wcsftime, but I'm quite concerned that it's failing in a way that I don't understand and can't reproduce.

[pganssle]

I'm able to reproduce this issue only on Ubuntu 14. For some reason with '\ud800', this function is not raising an error in Ubuntu 14. However, a simple POC:

#include <stdlib.h>
#include <stdio.h>

int main() {
    const wchar_t str[] = L"\xd800";
    char buffer[32];

    int rv = wcstombs(NULL, str, 0);

    printf("%d\n", rv);
}

Works fine on both Ubuntu 14 and other Linux (returns -1). I am honestly deeply puzzled at this. I also am not sure how to make a minimal triggering example, because I'm not sure what the pure python equivalent of PyUnicode_EncodeLocale is. '\ud800'.encode(errors="surrogateescape") doesn't hit this code path. Anyone have an idea?

[pganssle]

This is too annoying for me to continue working on it. Anyone else is free to pick up where I left off.