scArches scpoly fix

[aregano]

Following up some fixes I had to add for the HNOCA, I also needed to fix some tiny things on the scpoly function script to make it run properly. Fixes were:

1. Convert batch object to lumpy array, as it was built as a tensor and that would cause my pipeline to break
2. Same thing with the latent matrix
3. store x[batch, : ] as type torch.float32 to allow for gpu processing

You might have caught up to these issues, as I applied them during the spring, but you might not. Either way thanks for the neat integration models you provide in scArches!