Fstring and usage on TPU

Hi,

So I have been trying to make use of the composite optimizer in the tf-recommenders package for an object detection project. There are issues that result from training the model on the TPU, the issues mostly stem from the usage of python f-string when raising a warning or an error. These prevent TPU compilation without:

tf.config.set_soft_device_placement(True)

An example of this usage from the Composite Optimizer apply gradients method is the following:

raise ValueError(
    f"The set of variables handled by each optimizer should be "
    f"disjoint, but variable {v} is handled both "
    f"by {var_optimizer_dict[v.ref()]} and {optimizer}.")

This issue shows up in both tensorflow 2.4.0 and 2.5.0
NOTE: removing the f-string from the optimizer etc. allows for full usage on the TPU.