Fix TPU Convergence bug introduced by PR#6151 (#6488)

Currently with the bug introduced we're taking two optimizer steps per
batch: one global one, where `xm.optimizer_step` injects a CRS between
all cores in training, and one without. This has been affecting training
accuracy (for example, XLNet GLUE on MNLI is not converging, etc.).

src/transformers/trainer.py

@@ -572,7 +572,7 @@ class Trainer:
 
                     if is_torch_tpu_available():
                         xm.optimizer_step(self.optimizer)
-                    if self.args.fp16 and _use_native_amp:
+                    elif self.args.fp16 and _use_native_amp:
                         self.scaler.step(self.optimizer)
                         self.scaler.update()
                     else: