Use MultiHeadSelfAttention and MultiHeadGatedSelfAttention if want to return the attention weights

This commit adds new message-passing graph layers (MPN) and graph convolutional layers to dt, including vanilla MPN, GRUMPN, Masked-attention FGNN, and GraphTransformer.

Jp/rev magik

Fix indexing problems related to tf.gather

Allow multi-inputs in ContinuousGenerator

Fix phase correction at focus and offset calculation

 into develop

Add new graph datasets

Fix bug in the GRU gnn layer

Change loss input order of CGAN and PCGAN

Create detection_linking_hela dataset for GNN based models

CuPy fix for coherence length

Added a new method for testing MieSphere when coherence length parameter is provided.

Create node regression datasets for training GNN models

CuPy fix for MieScatterers with coherence length

Create virtual staining datasets and tools

removed pydeepimagej from deps

…into develop