Implement true CLT: per-source-layer encoders/decoders with causal masking

[bartbussmann]

Motivation

• Support a true Cross-Layer Transcoder (CLT) mode where features are per-source-layer and may write to multiple later target layers, and enforce causal source→target writes to match the CLT semantics.
• Make CLT configuration explicit by allowing multiple input layers (num_input_layers) so concatenated multi-layer inputs are handled correctly.
• Ensure training utilities (decoder normalization, auxiliary loss, dead-feature tracking, and BatchTopK) correctly handle the 3D activation tensors produced by true CLT mode.

Description

• Added num_input_layers to EncoderConfig and introduced multi_layer / true_clt flags in SharedEncoder to detect CLT mode via num_output_layers > 1 and num_input_layers > 1 and to compute input_size_per_layer when needed.
• Reworked encoder weights to support per-source-layer features: W_enc/b_enc become shape (num_input_layers, input_size_per_layer, dict_size) / (num_input_layers, dict_size) in true CLT mode, and added encode_linear to compute per-layer pre-activations.
• Expanded decoder to per-(source,target)-layer decoders W_dec with shape (num_input_layers, num_output_layers, dict_size, output_size) in true CLT mode and registered a causal_mask buffer so source layer s can only write to target layers t >= s, and applied that mask during decode and make_decoder_weights_and_grad_unit_norm.
• Updated update_inactive_features, auxiliary-loss path, BatchTopK flattening logic, and forward paths to correctly handle 3D activations ((B, S, D)) produced by CLT encoders and to preserve backward compatibility in non-CLT mode.
• Added CLT smoke test coverage: changed shape assertions for the concatenated-input CLT test and added test_clt_is_causal to verify perturbations to later-layer inputs do not affect earlier-layer outputs.

Testing

• Successfully compiled the modified modules with python -m py_compile base.py config.py test_clt.py which reported no syntax errors (success).
• Attempted to run python test_clt.py but the run failed in this environment with ModuleNotFoundError: No module named 'torch', so the full runtime tests could not be executed here (failure due to missing dependency).

---

Codex Task