Implement BitLinear Layer #178

Develop a core linear algebra routine for BitNet’s binary-weight linear layer (BitLinear). This function takes an input vector (or matrix) of 8-bit activations and multiplies it by a weight matrix with 1.58-bit (ternary) weights to produce an 8-bit output
ar5iv.org.

To do this efficiently, use bitwise operations: unpack the weight matrix bits (as prepared in step 3) on the fly, interpret 2-bit codes as -1, 0, or +1, and compute the dot-products. Exploit binary arithmetic where possible (e.g., treat -1 as +1 with a sign bit to use XNOR or bit-count tricks if inputs were binary, though here inputs are 8-bit). Accumulate results in a higher precision type (e.g. int32) then quantize/clamp the output back to int8 range. No bias terms should be added in these projections
ar5iv.org.

Make sure the BitLinear function is general and can handle weight matrices of various shapes (it will be used for Q, K, V projections, feed-forward layers, etc.). Optimize inner loops with bit-level operations and verify correctness on small test matrices.