math: finalize SLEEF audit and benchmark pass

[arduano]

Summary

• complete the full SLEEF revival audit pass and record keep/revert decisions in the math modules
• expand the benchmark harnesses so remaining f64 families expose forced-scalar and forced-backend evidence
• revert the losing f64 trig path to scalar-reference and keep the surviving portable/backend-specific wins

Validation

• cargo fmt --check
• cargo clippy --all --all-targets -- --deny warnings
• cargo test -q
• cargo test --release
• cargo clippy --all --all-targets --target aarch64-unknown-linux-gnu -- --deny warnings
• CARGO_TARGET_X86_64_UNKNOWN_LINUX_GNU_RUSTFLAGS="-C target-feature=+avx512f,+avx512bw,+avx512dq" cargo test --tests --target x86_64-unknown-linux-gnu --no-run
• CARGO_TARGET_X86_64_UNKNOWN_LINUX_GNU_RUSTFLAGS="-C target-feature=+avx512f,+avx512bw,+avx512dq" cargo test --release --lib --target x86_64-unknown-linux-gnu --no-run

Notes

• local CI parity uncovered and fixed one aarch64-only benchmark compile break in benches/simd_math_remaining_baseline/inverse_trig.rs
• local wasm and arm test execution were blocked by missing host tools on this machine, so GitHub Actions is the authoritative full-matrix confirmation