Patch ctgan batchsize1 (#263)

* reuse encoders

* ensure categorical encoder is trained on real and synthetic

* better transformer

* remove unnecessary imports

* better error message

* compatbility with DDIM

* Fix CTGAN for batchsize=1

A .squeeze statement caused errors during training when one of the batches is 1. Ensuring only the last dimension is squeezed fixes this

src/synthcity/plugins/core/models/gan.py

@@ -682,7 +682,7 @@ def _loss_gradient_penalty(
         interpolated = (
             alpha * real_samples + ((1 - alpha) * fake_samples)
         ).requires_grad_(True)
-        d_interpolated = self.discriminator(interpolated).squeeze()
+        d_interpolated = self.discriminator(interpolated).squeeze(-1)
         labels = torch.ones((len(interpolated),), device=self.device)
 
         # Get gradient w.r.t. interpolates