Tracing JIT: give the type guard on entry of an inlined call an exit opline

[waroad]

Here is the patch I offered in the issue, thanks for the go-ahead.

When the tracing JIT inlines a call (ZEND_JIT_TRACE_ENTER) and some of the callee's CVs are register-allocated with ZREG_LOAD and still MAY_BE_GUARD, it emits a type guard for them with a NULL opline. A NULL exit opline means "EX(opline) is already valid", but at this point the frame is the freshly pushed callee whose EX(opline) has not been written yet. So when the guard fails (here: the reference-bound arguments are ints while the trace is recorded and floats on a later call), zend_jit_trace_exit() resumes the VM at a garbage opline. A debug build trips the opline assertion; a release build segfaults (as reported in the issue).

The fix passes (p+1)->opline, the callee's first traced opline, the same opline the SEND_UNPACK/SEND_ARRAY guard just above uses. A failing guard then exits to the start of the callee instead.

The test ext/opcache/tests/jit/gh24389.phpt is the reproducer from the issue, plus a check of the return values. On a PHP-8.4 debug ASan build it fails the zend_jit_trace_exit assertion without the fix and passes with it. The issue's reproducer and its foreach by-reference variant now run to completion, and the variant prints the same as with the JIT off. With the fix, ext/opcache/tests has the same three failures under the tracing and the function JIT (gh9259_003, preload_parse_error, log_verbosity_bug). They are LeakSanitizer reports from this ASan build, and they fail the same way without the fix under the tracing JIT.

Fixes #24389.

Found with an LLM-based testing tool; I wrote this fix with help from Claude and reviewed it myself.