[fhgffy]

Fixes #502.

glm_spin currently right-multiplies a rotation built around the matrix's translation. Translating to (5, 0, 0) and spinning 90 degrees about Z changes the position to (10, -5, 0).

Build an origin-centered rotation instead. For an affine matrix [A, t; 0, 1], the existing right multiplication then produces [A R, t; 0, 1], preserving translation and the pre-transform multiplication order. The shared implementation also fixes glmc_spin and glms_spin.

Add five registered regressions covering translated matrices, an existing rotation with nonuniform scale, a nonunit axis, inverse and zero angles, rotation at the origin, and struct input preservation. Expected matrices are written explicitly.

Validation on the unchanged patch:

• Original Linux suite: 1131/1131. With only the new tests, all five regressions fail while the original 1131 pass.
• Linux GCC 13.3, C99: 1136/1136; also 1136/1136 with AddressSanitizer and UndefinedBehaviorSanitizer.
• Windows MinGW GCC 15.1, complete official suite compiled directly with GCC: 1136/1136. Focused C99 checks with -Wconversion -Werror: 9/9.
• Additional double-precision Rodrigues checks on both Windows and Linux: 515 affine cases, three APIs, 4635 matrix comparisons per platform, zero failures. The Linux run also uses ASan/UBSan.

The Windows CMake MinGW generator failed before compilation in the Unicode workspace path; its build is not claimed as passing. The Linux C99 suite retains pre-existing anonymous-struct warnings. MSVC, macOS, ARM/NEON, WebAssembly, the full SIMD build matrix, and upstream CI have not been run.