Implement the 128-bit rounded store

[NiceAndPeter]

Implements the roadmap item added in #16: when the cold assignment path's exact 64-bit formation of (rhs − Lower)/Notch overflows (full-mantissa fp-derived sources on grids with large |Lower|, e.g. math::dbl::tan's auto output grid), the offset slot is now computed directly in wide arithmetic instead of reporting errc::overflow.

How

• rational.hpp: mul128 (128×64 product with overflow flag) and divmod128 (128÷64 quotient+remainder; native __int128 or the portable shift-subtract construction already used by to_fixed's fallback — MSVC covered).
• assignment.hpp: wide_offset_quotient computes (rhs − Lower)·d_n / (a_dr·a_dl·n_n) with the compile-time divisor factors gcd-reduced first, so the only wide operations are one 128×64 multiply and one 128÷64 divide. store_checked applies the normal policy ladder (exact store / round / rounding_error) to the result.
• generic.hpp: offset rounding rules factored into round_offset, shared by round_quotient and the wide store — rounding semantics in the >64-bit regime are identical everywhere (offset rule; differs from value-space only at exact ties on negatives).

Only results beyond even the 128-bit envelope still report errc::overflow.

Caveat (documented in roadmap.md): at constant evaluation the transient rational overflow surfaces as the intentional constexpr_error diagnostic before the fallback can engage — wide-regime stores are runtime-only in practice. Changing that would mean revisiting the consteval overflow convention library-wide.

Tests

• The wide-grid store regression now asserts the positive case stores the correctly rounded slot (previously pinned as errc::overflow), plus a strict-policy rounding_error case.
• New test_cmath_double.cpp case: dbl::tan at full-mantissa points on an integer-index snap grid.
• The original 49k-point tan sweep repro runs clean end to end.

Verification

429/429 ctest (unit, examples, compile-fail, amalgamate-up-to-date, codegen guard) · 418/418 under BND_MATH_FIXED · fuzz seed 7 × 20k iters clean (110.6M property checks) · clang-tidy clean under the CI header filter · amalgamation regenerated.

🤖 Generated with Claude Code