Update trainer.py

Cleaning each time you compile the model the previous trainer metrics also erased the compilation of other models with are a cominbation of that model. Given problems in archiquectures shuch as GAN. As explain in issue:

keras/src/trainers/trainer.py

@@ -140,7 +140,6 @@ def compile(
                 wrapped in a `LossScaleOptimizer`, which will dynamically
                 scale the loss to prevent underflow.
         """
-        self._clear_previous_trainer_metrics()
         optimizer = optimizers.get(optimizer)
         self.optimizer = optimizer
         if (