Add OpenBLAS dense backend to linalg

[MesTTo]

Adds an optional OpenBLAS-backed dense f32 backend to the linalg crate, behind the existing blas feature.

blas_backend exposes row-major dense matmul and attention (scores + apply) kernels that delegate to OpenBLAS sgemm, for standard dense workloads where a vendor BLAS beats a hand-written kernel. The module is gated behind blas, which now also enables dense (the backend builds on Dense). No default build changes: blas stays opt-in.

The tests check the kernels against hand-computed results:

• rectangular matmul matches the expected product
• attention scores match manual dot products
• attention apply matches a manual weighted sum

Validation (OpenBLAS installed):

• CARGO_HOME=/home/user/Dev/.cargo-isolated RUSTFLAGS='-C target-cpu=native -Awarnings' cargo +nightly test -p linalg --features blas blas_backend
• rustfmt +nightly --edition 2024 --check linalg/src/blas_backend.rs
• git diff --check
• jscpd --reporters ai --min-lines 5 --min-tokens 50 linalg/src/blas_backend.rs