tensorflow: tf.function and tf.nest break for valid Mapping instances

tf.function makes invalid assumptions about arguments that are Mapping instances. In general, there are no requirements for Mapping instances to have constructors that accept [(key, value)] initializers, as assumed here.

This leads to cryptic exceptions when used with perfectly valid Mapping subclasses such as this one:

class CustomMapping(Mapping):

  def __init__(self, **kwargs):
    self.mapping = kwargs
  
  def __getitem__(self, key):
    return self.mapping[key]
  
  def __iter__(self):
    return iter(self.mapping)

  def __len__(self):
    return len(self.mapping)

See this Colab notebook for an example.