You could probably use that formula to combine standard normal and chi-sq samples into nct samples. But to run inference you need the formula for the logpdf of the distribution.
This link may also be useful to you. The code looks quite challenging to port into Theano to be honest.
You can pass numpy arrays into PyMC3 models as inputs. You should be able to use them for static data within Theano functions. But where your input is a theano variable and your output is a theano variable then you shouldn’t use numpy. Say if you want to use a PyMC3 distribution as an input.
Hope this is helpful, sorry I cannot be more useful