[codex] Enable contiguous mul range path without parallel

[shinaoka]

Summary

• compile the contiguous-range multiply planner and runner without the parallel feature
• keep the Rayon split path behind parallel, while sharing the single-thread runner for both builds
• add a regression test that the high-rank contiguous broadcast-mul plan is available in the default non-parallel build

Root cause

Tenferro uses strided-kernel without the parallel feature in the benchmark path. The large contiguous-range broadcast-multiply path was gated entirely by parallel, so 1T runs fell back to the generic strided loop for large intermediates.

Validation

• cargo test -p strided-kernel -- --nocapture
• cargo test -p strided-kernel --features parallel -- --nocapture
• cargo fmt --check && git diff --check