port the lens read to llama.cpp

[solarkyle]

the lens is just per-layer matmuls on hidden states. llama.cpp has tensor callbacks (cb_eval). prototype exists in pytorch here, someone who knows the ggml side could make the router usable where people actually run models