Expand strided einsum CPU kernels

[shinaoka]

Summary

• move broadcast multiply / batched outer product / non-conjugated dot-general execution into reusable strided-rs kernels
• add strided-einsum2::DotGeneralConfig with borrowed axis slices and BLAS provider feature aliases
• add PyTorch comparison benches and kernel-writing notes for future SIMD/kernel work

Why

tenferro-rs now delegates CPU binary einsum kernels to strided-rs, so these paths need to live behind general stride/layout APIs rather than benchmark-specific tenferro shortcuts.

Notes

• The implementation is stride/layout driven; it does not branch on benchmark names or fixed tensor shapes.
• docs/superpowers/... local planning notes were intentionally not committed.

Verification

• cargo fmt --check
• cargo test -p strided-kernel --features parallel
• cargo test -p strided-einsum2 --features parallel,blas-accelerate --no-default-features
• git diff --check