prevent deadlock when releasing the jl_unique_gcsafe_lock causes gc (#…

…56563)

Caught this by running threads test repeatedly locally: the sweep needs
to acquire engine_lock, so we need to make sure to release that first
(the other jl_unique_gcsafe_lock users shouldn't care about this
ordering since they don't acquire their locks during sweeping)

src/julia_locks.h

@@ -115,15 +115,16 @@ class jl_unique_gcsafe_lock {
     explicit jl_unique_gcsafe_lock(std::mutex &native) JL_NOTSAFEPOINT_ENTER
     {
         jl_task_t *ct = jl_current_task;
-        gc_state = jl_gc_safe_enter(ct->ptls);
+        gc_state = jl_gc_safe_enter(ct->ptls); // contains jl_gc_safepoint after enter
         this->native = std::unique_lock(native);
         ct->ptls->engine_nqueued++; // disables finalizers until inference is finished on this method graph
     }
     jl_unique_gcsafe_lock(jl_unique_gcsafe_lock &&native) = delete;
     jl_unique_gcsafe_lock(jl_unique_gcsafe_lock &native) = delete;
     ~jl_unique_gcsafe_lock() JL_NOTSAFEPOINT_LEAVE {
         jl_task_t *ct = jl_current_task;
-        jl_gc_safe_leave(ct->ptls, gc_state);
+        native.unlock();
+        jl_gc_safe_leave(ct->ptls, gc_state); // contains jl_gc_safepoint after leave
         ct->ptls->engine_nqueued--; // enable finalizers (but don't run them until the next gc)
     }
     void wait(std::condition_variable& cond) JL_NOTSAFEPOINT {