DestVI log-variational for Vamp_prior

[canergen]

trafficstars

In models/_condscvi.py line 139, we don't check for log_variational (default is true in CondSCVI). The input for the decoder at inference is therefore different from the training input. Effect is not clear, but most likely unwanted.

for tensors in scdl:
    x = tensors[REGISTRY_KEYS.X_KEY]
    y = tensors[REGISTRY_KEYS.LABELS_KEY]
    out = self.module.inference(x, y)

Fix is straightforward but leads to breaking change. It should be similar to model.get_latent_representation() meaning:

inference_inputs = self.module._get_inference_input(tensors)
out = self.module.inference(**inference_inputs)