gc: prevent LTO from eliminating rb_gc_before_fork

GCC LTO with -flto=auto creates a constprop clone of rb_fork_ruby where
rb_multi_ractor_p() is constant false. In that clone, rb_gc_vm_barrier()
and RB_GC_VM_LOCK() are both no-ops, so rb_gc_impl_before_fork reduces
to a trivial store and GCC eliminates the entire call chain. This leaves
no GC barrier before fork(2).

RBIMPL_ATTR_NOINLINE() on rb_gc_before_fork prevents GCC from inlining
it into the constprop clone, forcing it to treat the call as having
unknown side effects.

Additionally, call gc_rest() at the top of rb_gc_impl_before_fork to
complete any in-progress incremental GC cycle before forking. This
ensures the child does not inherit a mid-cycle heap.

Author: ianks

gc.c

@@ -5435,6 +5435,7 @@ rb_obj_info_dump_loc(VALUE obj, const char *file, int line, const char *func)
     fprintf(stderr, "<OBJ_INFO:%s@%s:%d> %s\n", func, file, line, rb_raw_obj_info(buff, 0x100, obj));
 }
 
+RBIMPL_ATTR_NOINLINE()
 void
 rb_gc_before_fork(void)
 {

gc/default/default.c

@@ -9407,6 +9407,8 @@ rb_gc_impl_before_fork(void *objspace_ptr)
 {
     rb_objspace_t *objspace = objspace_ptr;
 
+    gc_rest(objspace);
+
     objspace->fork_vm_lock_lev = RB_GC_VM_LOCK();
     rb_gc_vm_barrier();
 }