feat: add MoLT (Mixture of Linear Transforms) model

[Ky-Ng]

Port the MoLT model from the original pre-reorg branch onto the current crosslayer_transcoder/ package layout.

• New crosslayer_transcoder/model/molt.py containing the Molt nn.Module (paired U/V factors across a configurable list of ranks, a shared gate produced by an arbitrary nonlinearity, and a transform_norm() helper for weighted sparsity).
• New MoltModule in crosslayer_transcoder/model/clt_lightning.py built on top of CrossLayerTranscoderModule: single-layer training loop, sparsity penalty on gate * ||U@V||, and forward override.
• CrossLayerTranscoderModule.init accepts Union[CrossLayerTranscoder, Molt] and sizes last_active from model.n_features when the model is a Molt.
• JumpReLU supports n_layers=1 by dropping the layer dimension from theta (shape becomes (1, d_features)). n_layers > 1 is unchanged.
• config/molt.yaml with class_paths wired to the current package.