Text this: Addressing Distribution Discrepancies in Pulsar Candidate Identification via Bayesian-neural-network-based Multimodal Incremental Learning