Fix unsound numeric_minimum bounds and add symmetric numeric_maximum

[gl84]

numeric_minimum relied solely on is_numeric::value to decide whether to query std::numeric_limits. For types where is_numeric is true but numeric_limits is not specialized (e.g. boost::rational), min()/max() silently fall back to a default-constructed T(), producing wrong bound checks (#58). It also used min() for the lower extreme, which for non-integral types is the smallest positive value rather than the most negative one.

Changes:

• Guard all bound checks on std::numeric_limits::is_specialized; unspecialized numeric domains now impose no constraint instead of comparing against a bogus default-constructed bound.
• Use lowest() (not min()) for the lower extreme.
• Add a numeric_maximum trait mirroring numeric_minimum to detect domain_next overflow, and assert it wherever an interval boundary is built via domain_next (singleton, unit_trail, hull) and in first(), making first()/last() symmetric.
• Add regression tests for the bound traits (int, double, rational) and for first()/last() across all bound flavors.

Fixes #58