Consider binarizing the weights in the splat decoder

Here, we have a scalar value which adjusts the weight of each Gaussian in the decoder:

https://github.com/alan-turing-institute/affinity-vae/blob/23349e544dc86cf335ab0e948ccea333f40ca3b4/avae/decoders/differentiable.py#L188-L195

We could consider using [binarization](https://proceedings.neurips.cc/paper_files/paper/2016/file/d8330f857a17c53d217014ee776bfd50-Paper.pdf) of this linear layer to improve the reconstruction.

Start with avae/decoders/differentiable.py around lines 188-195 and read the linked binarization paper. Determine how the decoder's scalar Gaussian weights and reconstruction are evaluated before defining the binarized-layer experiment. Done means implementing the agreed approach and demonstrating improved reconstruction without breaking decoder behavior.

Written by the indexing model from the issue text.