feat(ops): add LinearBitNet — ternary weight GEMV with zero-skip

[eriirfos-eng]

What this adds

A LinearBitNet struct in crates/ruvector-sparse-inference/src/ops.rs, alongside the existing Linear.

The difference: weights are stored as i8 in {−1, 0, +1} rather than f32. The forward pass skips every multiply-accumulate where the weight is zero — no approximation, no special hardware. At the sparsity levels typical of BitNet b1.58 and similar ternary-quantized models (50–70% zeros), this halves the active MACs in the linear layer.

Why it belongs here

ruvector-sparse-inference already has a SparseFfn that exploits activation sparsity (skipping neurons that activate to zero). LinearBitNet is the complementary case: weight sparsity, where the zeros are baked into the model at quantization time rather than determined at runtime.

API

// Quantize an f32 weight matrix at load time
let layer = LinearBitNet::from_f32(out_features, in_features, &weights_f32, threshold, bias);

// Sparse forward — zero weights are skipped in the inner loop
let output = layer.forward(&input);

// Fraction of weights that are zero (useful for benchmarking)
let s = layer.sparsity();

threshold is the absolute value below which a weight becomes 0. The BitNet b1.58 paper uses the mean absolute value of the weight matrix as the threshold, which works well in practice.

Changes

• crates/ruvector-sparse-inference/src/ops.rs: LinearBitNet struct with from_f32, forward, sparsity
• Three tests: forward correctness (hand-verifiable), sparsity reporting, sign-consistency with dense at zero threshold

What it does not change

Existing Linear, Embedding, RMSNorm, LayerNorm and all existing tests are untouched. All 91 existing tests pass.