fix data modification for tensorflow compiled mode

bayesflow/experimental/backend_approximators/base_approximator.py

@@ -44,6 +44,10 @@ def evaluate(self, *args, **kwargs):
 
  # noinspection PyMethodOverriding
  def compute_metrics(self, data: dict[str, Tensor], stage: str = "training") -> dict[str, Tensor]:
+ # compiled modes do not allow in-place operations on the data object
+ # we perform a shallow copy here, which is cheap
+ data = data.copy()
+
  if self.summary_network is None:
  data["inference_variables"] = self.configurator.configure_inference_variables(data)
  data["inference_conditions"] = self.configurator.configure_inference_conditions(data)