Infer dim_input in tf_model_example

[fhoekstra]

I found this fixed an error that was otherwise raised by get_input_layer in my attempt at a box2d_arm_img experiment.

I am not sure where in the initialization process of the experiment the dim_input is passed along, but found out that it was zero here. Adding this 1 line per multi_modal_network fixes the issue by inferring dim_input from the given obs_include hyperparameter.