In-band full duplex-based cell-free (IBFD-CF) systems suffer from severe interference problem including self-interference (SI) and cross-link interference (CLI), especially when cell-free (CF) systems are operated in a distributed way. To this end, we propose multicarrier-division duplex (MDD) as an enabler for full-duplex (FD)-style operation in distributed CF massive MIMO systems, where DL and UL transmissions take place simultaneously at the same frequency band but mutually orthogonal subcarrier sets. In order to maximize the spectral efficiency (SE) in the proposed systems, we present heterogeneous graph neural network specific for CF systems (CF-HGNN), which consists of an adaptive node embedding layer, meta-path based message passing, meta-path based attention and downstream power allocation learning. In particular, the adaptive node embedding layer can handle the varying number of access points (APs), mobile stations (MSs) and subcarriers, and the involved attention mechanism enables each AP/MS node in CF-HGNN to aggregate the information from interfering path and communication path with different priorities. Numerical results show that CF-HGNN is capable of using $10^4$ times less operation time to achieve the 99% performance of the SE of quadratic transform and successive convex approximation (QT-SCA). Additionally, CF-HGNN also significantly outperforms unfair greedy method in terms of SE performance. Furthermore, CF-HGNN exhibits good adaptivity to varying number of nodes and subcarriers, and also generalization ability to different sizes of CF network.