Hardware-efficient tractable probabilistic inference for TinyML Neurosymbolic AI applications