You might find the implementation from @AustinRochford helpful PyMC3 BSplines · GitHub
For your second question, both solutions are not recommended. If you are observing problem in the MCMC run, such as sampler being stuck, divergent samples, it is an indication that there are some hidden problem of your model: some area of the parameter space contains high curvature, etc.
I would suggest first try more informative priors to cut out those part of the parameter space, then explore different parameterization of your model.