* added optimized paths for matrix-vector and vector-matrix products
  (using either a cache friendly strategy or re-using dot-product
  vectorized implementation)
* add LinearAccessBit to Transpose