fix: Allow hooks to be removed after class binding

[PROFeNoM]

Description

During the micro benchmarking process, I came across an unexpected behavior related to #2367. The goal of the latter was most notably to introduce minimal overhead when injecting logs while instrumenting Psr\Log\NullLogger. However, the results showed otherwise (refer to the Before/After comparison below).

The problem arises when the call to \DDTrace\remove_hook is made after a class bind (zen_compile.c:do_bind_class). As a result, the observers-related code (zend_observer.c: _zend_observer_class_linked_notify) does not get executed. This implies that hook.c:zai_hook_resolve_class and hook.c:zai_hook_is_excluded are not called either. Thus, when the supposedly removed class is called again, there is no check to see if the hook was removed or not and we directly go to zend_observer.c:_zend_observe_fcall_begin & hook.c:zai_hook_continue. This pull request addresses this issue by adding the necessary check.

From a general perspective, it was only effective to remove a hook before a class binding occurs. Attempting to remove a hook at any other point was ineffective.

While microbenchmarks are still being developed, the before/after comparison below shows the effect of logs injection (Injection) versus no logs injection (Baseline):
Before:

After:

Reviewer checklist

• Test coverage seems ok.
• Appropriate labels assigned.

[realFlowControl]

The tests are looking good and are ✅
@morrisonlevi knows more about the tracer, so if it is okay, let's wait for his review

[morrisonlevi]

I'm not sure about the intricacies about removing hooks and such. I've reviewed code around this before but it's been a while and nothing is fresh. It would take quite some time for me to get back into it. It looks fine to me, but if you want to be extra careful and have me take the extra time to review the deeper impact, let me know.

[bwoebi]

Superseded by #2449.