Fix GCC optimization bug eliminating rb_gc_before_fork()

GCC 15.2.0 with -O3 and -fipa-cp-clone completely eliminates
rb_gc_before_fork() because it sees rb_gc_impl_before_fork() is a no-op.
This removes the memory synchronization point before fork(), causing
stale cached heap slot data to persist in child processes, leading to
jemalloc corruption and SIGSEGV crashes.

Add __atomic_thread_fence(RBIMPL_ATOMIC_SEQ_CST) to force GCC to preserve
the function and create a proper memory barrier. Uses Ruby's memory
ordering constant from include/ruby/atomic.h. This generates a
lock-prefixed instruction on x86-64, preventing function elimination
and ensuring memory synchronization across fork boundaries.

Affects: Shopify's +optimized builds with GCC 15, -O3, --disable-shared
Impact: Fixes intermittent crashes in Puma/Unicorn worker processes

Author: ianks

gc.c

@@ -5618,6 +5618,7 @@ void
 rb_gc_before_fork(void)
 {
     rb_gc_impl_before_fork(rb_gc_get_objspace());
+    __atomic_thread_fence(RBIMPL_ATOMIC_SEQ_CST);
 }
 
 void