I would recommend going through http://deeplearning.net/software/theano/tutorial/adding.html to get a sense of theano variables (the same kinds of principles will apply to tensorflow and pymc4).
In brief, theano variables are not assigned a value until you pass them in explicitly. shared objects are the exception to this rule, and are instantiated with values at the outset.
The reason you obtain the same results is far simpler: if you pass in a numpy array as observed data, pymc3 applies theano.shared as a convenience.