Optimize strided permutation fast paths

[shinaoka]

Summary

• add tiled transpose-scale fast paths for f64 and identity-copy scalar types, including zero-scale handling
• add HPTT-style grouped high-rank permutation planning and parallel execution with a serial fallback when Rayon has one worker
• add tile-level parallel execution for rootless 2D transpose plans, fixing the 3D [2,0,1] parallel scaling bottleneck
• tighten strided-perm public API and move benchmark coverage out of the crate README into the benchmark suite
• add repository rules adapted from tenferro-rs, including the nthreads=1 serial-kernel rule
• link the top README to the benchmark suite and invite bottleneck/failure-case reports

Verification

• cargo fmt --check && cargo test -p strided-perm --features parallel && cargo test
• cargo llvm-cov --workspace --json --output-path coverage.json && python3 scripts/check-coverage.py coverage.json

Benchmarks

M5 MacBook Pro numbers are recorded in tensor4all/strided-rs-benchmark-suite on branch scale-transpose-bench. The 3D 256^3 [2,0,1] parallel path improved from about 15.8 ms to 4.792 ms on 4 threads.