fix: improve efficiency and accuracy of value_cast for quantity_point

[mpusz]

Overview

Rewrites the quantity_point overload of detail::sudo_cast to select the optimal intermediate unit and representation type for each conversion case, avoiding unnecessary operations and improving numerical accuracy.

This implements the design from #598 (originally authored by @burnpanck). The commit carries a Co-authored-by trailer to credit the original work. PR #598 has been closed in favour of this one.

Problem

The previous implementation always delegated cross-origin conversions through a recursive point_for() call (twice for the small-unit case), unconditionally rescaling to a common type before adding the origin offset. This forces extra arithmetic operations in cases where none are needed, and in the floating-point path could accumulate rounding error by expressing the offset in the wrong unit.

Additionally, there was a latent bug in the same-origin branch: the returned quantity_point was tagged with FromQP::point_origin instead of ToQP::point_origin.

Solution — four branches

Condition	Strategy
Same origin	Delegate entirely to the quantity sudo_cast. Also fixes the origin tag bug.
Same unit, different origin	Skip unit scaling; compute the static origin offset in the common rep and add it directly. Always exact for integers.
Different unit, floating-point rep	Pick the larger unit as the intermediate (minimises scaling magnitude), then call point_for() to let the library's common-unit arithmetic add the offset. This avoids precision loss that would occur if the offset were expressed explicitly in the larger unit (e.g. 42 m → 0.042 km loses bits).
Different unit, integer rep	Prefer the output unit as intermediate (result is already in the right unit before truncation). Fall back to the input unit if the origin offset in the output unit would overflow the common rep type.

Tests

The existing value_cast tests for quantity_point are expanded to exercise every branch:

• Branch 1 (same origin): existing tests retained.
• Branch 2 (same unit, different origin): integer rep (both directions), and fractional double.
• Branch 3 (FP, unit change): existing km→m test (from-unit larger); new m→km test (to-unit larger).
• Branch 4 (integer, offset fits in output unit): existing int8_t mm → cm pair.
• Branch 5 (integer, offset overflows output unit): new int16_t m → mm test where the 42 000 mm offset overflows int16_t, so the input unit (m) is used instead.

Related

Closes #598