core/tensor/blas library
OpenBLAS FFI bindings for the CPU matmul fast path.
Blas.sgemm wraps cblas_sgemm from the system-installed OpenBLAS
(libopenblas.so.0). Falls back gracefully when the library isn't
present or the caller sets DART_PYTORCH_NO_BLAS=1 — in either
case Blas.enabled returns false and callers should use the
Dart matmul path.
On WSL 2 with libopenblas-pthread 0.3.26, this gives roughly a
20-50× speedup on the shapes we hit in Qwen1.5-MoE forward
([T, D=2048] @ [D, hidden=1408..151936]) compared to the
hand-unrolled Dart triple loop.