feat(hslm): sparse ternary matmul — skip zero weights

[gHashTag]

Summary

Sparse ternary matrix multiplication that only iterates non-zero entries.

New file

• src/b2t/sparse_ternary.zig — 203 LOC

Performance

• Dense matmul: O(rows × cols) — iterates all entries including zeros
• Sparse matmul: O(nnz) — only non-zero entries stored and processed
• For typical ternary sparsity (~33% zeros): ~3x fewer operations

Features

• SparseTernaryMatrix: CSR-like sparse storage, init from dense weight array
• matmul(): O(nnz) sparse matrix-vector multiply
• matmulBatch(): Batched version for sequence processing
• SparseStats: nnz count, sparsity %, compression ratio
• computeStats(): Statistical summary of sparsity

Tests (5)

• Sparse output matches dense reference
• All-zero weights → nnz=0, sparsity=100%
• All-nonzero weights → nnz=total, sparsity=0%
• Sparse stats accuracy
• Batch matmul correctness

Closes #316