Fix `build_keras_loss` util

With the release of tensorflow 2.13.0, tf.keras.losses.deserialize no longer raises a ValueError when invalid string inputs are provided. As a result, calling declearn.model.tensorflow.build_keras_loss("mse") results in a TypeError when it should (and used to) return a custom tf.keras.losses.Loss instance wrapping the tf.keras.losses.mse loss function.

An additional task that should be performed as part of tackling this issue is implementing unit tests for this util (as the bug only emerged when running integration tests).