Fix runtime errors in vbgan.py

[lxt3]

Proposed fix for: RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [128]] is at version 3; expected version 2 instead

1. Separate forward passes for the encoder and decoder updates
2. Freeze encoder and use x.detach() for decoder pass: After stepping the encoder, its batch-norm statistics are updated. By freezing encoder parameters, detaching the input, and recomputing the forward pass, the decoder receives a fresh computation graph that ignores the encoder’s updated state. This avoids the “in-place modification” runtime error.
3. No retain_graph: Because each backward now uses its own fresh forward pass, the graph is rebuilt every time. No shared graph means no need to retain it

Also, pass in empty list to parse_args(): args = parser.parse_args([])