Define Data Structures & Load Weights #173

Design Go structures to hold the model parameters (e.g. a struct for each Transformer block containing its weight tensors). Implement a loader that reads the embedded weight file and populates these structures. This will involve unpacking the 2-bit ternary weights (-1 → 00, 0 → 01, +1 → 10; 11 unused): the model uses ternary weight values {-1, 0, +1}, which are packed as four 2-bit values per byte
ar5iv.org.

Decode each byte by extracting 2-bit segments and mapping them to -1/0/+1 (using the quantization scheme from training that maps weights to ternary values
ar5iv.org
).

Verify that the shapes of the loaded weight matrices match the expected dimensions of the model layers.

Loader must also read the per-matrix abs-mean float32 scale that accompanies every packed tensor and store it alongside the weights for later de-quantization.