COMP: Skip fpclassify for integral vnl_math classification predicates

[hjmjohnson]

Skip fpclassify for integral vnl_math classification predicates so integer-typed vnl_matrix/vnl_vector compile on MSVC toolsets where fpclassify(<integral>) is ambiguous.

Root cause

numeric_predicates::isfinite/isnan/isinf/isnormal forward std::isfinite(arg) (etc.) for every TArg. For integral TArg this instantiates the Windows UCRT <corecrt_math.h> integer template

template <class _Ty> inline bool isfinite(_In_ _Ty _X) throw() { return fpclassify(_X) <= 0; }

where fpclassify(<integral>) is an ambiguous call (error C2668: float/double/long double) on e.g. MSVC 14.38.33130 (cl 19.38) + Windows SDK ≤ 10.0.22621. Every integer-typed vnl_matrix/vnl_vector TU fails to compile; newer toolsets (MSVC 14.44) are unaffected.

Fix: short-circuit integral arguments with if constexpr — an integer is always finite, never inf/NaN, normal when non-zero — avoiding the ambiguous SDK template.

Validation

The equivalent change applied to ITK's vendored copy of this file builds all of ITK (3396/3396 targets, 0 errors) on the failing toolchain (MSVC 14.38.33130, Windows SDK 10.0.22621.0). Carried as ITK PR InsightSoftwareConsortium/ITK#6451 pending this upstream fix.

AI assistance

• Tool: Claude Code (claude-opus-4-8)
• Role: diagnosed the ambiguous SDK fpclassify integer template, authored the if constexpr guard, validated via a full ITK build.

[hjmjohnson]

Downstream consumer: InsightSoftwareConsortium/ITK#6451 carries this same change as an ITK-local patch to its vendored vnl_math.h, to unblock the ITK build until a VXL sync pulls in this fix. That ITK-local patch can be removed once this merges and is synced into ITK.