use_session_with_seed broken with TensorFlow 2.0

[apiskors]

The R function tensorflow::use_session_with_seed calls tf$ functions four different times. With the Python TensorFlow 2.0 libraries, every one of those tf$ calls fails, e.g. like so:

Error in py_get_attr_impl(x, name, silent) : AttributeError: 'module' object has no attribute 'reset_default_graph'

It is possible to make the errors go away by replacing all occurrences of tf$ in the source code with tf$compat$v1$, but I suspect that may not actually give the correct functionality.

[apiskors]

I found part of the same problem reported as keras issue #890.