Function std::tensor_matmul_auto

&FloatTensor tensor_matmul_auto(const &FloatTensor a, const &FloatTensor b)

No size check at all: benchmarked 2x2x2 through 128x128x128 (8 to ~2,097,152 multiply-adds) and BLAS won or tied at every single point — no naive-wins region was ever found, unlike vDSP's real one above. This project's own two training-loop benchmarks (128x256x64 and 128x512x1024x256) independently confirm the same thing at much larger scale. Always dispatching to BLAS here isn't a shortcut that skipped measuring the small-size case — it's what measuring the small-size case actually found.