[Observer] Fatal error support

[SammyK]

This is another follow up PR to #5857.

This fixes an issue where fatal errors would not call the observer end handlers. And also adds a few missing tests.

[morrisonlevi]

Looks good to me.

[nikic]

Does this work with ext/soap or other exts that throw fatal errors and then catch the bailout?

[beberlei]

@nikic Yes it should be, zend_error_impl is called regardless of being "in soap" or not for errors/warnings/... when they occur, this happens outside the php_error_cb which can be replaced, its placed near the error notify callback code, which was also built to always trigger.

[nikic]

@beberlei My concern is the other way around: The fatal error will be caught by soap and execution will continue, exiting some frames that have already been exited.

[SammyK]

@nikic Very good point. I can confirm that the end handlers get fired twice in the case of ext/soap fatal errors. I'll see if I can come up with an elegant solution to handle the edge cases that extensions like ext/soap present. 🤔

Tangent: Ideally this PR would address all zend_bailouts at a general level, but I'm pretty sure addressing all the edge cases is impossible. I'm wondering if php-src is at the point now where we can seriously discuss a plan to kill zend_bailout with fire? 🔥

[SammyK]

Here's a possible solution to extensions that juggle errors like ext/soap: 687b7c1.

The idea is that for fatal errors, functions are "unobserved" after the end handlers fire for the first time. This causes all the observer handlers to fire and close prematurely for these specific edge cases, but ensures they only fire once.

The long term solution would be: fix each individual edge case of fatal error juggling.

[SammyK]

@nikic I've looked at this a lot of different ways to see if we can get something akin to ZEND_HANDLE_EXCEPTION for remaining stack frames on shutdown after a zend_bailout. I don't think there is a safe way to traverse the stack on shutdown which makes me think we're stuck with the above solution of a one-time call to the end handlers on fatal errors.

I'm open to other ideas of addressing zend_bailout (ab)use. /cc @morrisonlevi @beberlei

[SammyK]

Closing this in favor of #6377 which provides a more generic solution for handling zend_bailout.