Yeah, not having BLAS will definitely slow things down, though GPs do tend to be a little slow (in your case, inverting a 400x400 matrix). I usually have luck installing via pip instead of conda when this happens.
Definitely use either version 4 or 5 of PyMC, rather than PyMC3 as Christian suggests.
Also, you may want to try using pymc.sampling_jax.sample_numpyro_nuts instead of pymc.sample – it will generally be faster.