While your question is from last year, I still wanted to chime in to maybe provide some help.
If you check the answer at Passing hidden layers to ConvLSTM, I think you should be able to modify the code to utilize Conv3d layers instead of their 2D counterparts.
Cheers,
Petteri