Preserve cached rotations during sprite collision queries

[fderop]

Sprite collision queries copied rotation matrices but discarded their cached coefficients, so later transforms repeated sine and cosine calculations.
The copy now preserves the coefficients and their validity flag:

m_Elements[0][0] = reference.m_Elements[0][0];
m_Elements[0][1] = reference.m_Elements[0][1];
m_Elements[1][0] = reference.m_Elements[1][0];
m_Elements[1][1] = reference.m_Elements[1][1];
m_ElementsUpdated = reference.m_ElementsUpdated;

Valid caches remain valid. Invalid caches still recalculate on use.
MOSprite::HitTestAtPixel also drops its redundant matrix copy. The existing const operator retains a local copy, so queries do not mutate shared matrices.

Three process pairs on an M3 Pro with GCC 15.2 show 10.7–24.6% less collision-query time, depending on cache state.
All 29,491,200 pixel-query results and matrix float-bit hashes match. Concurrent-reader checks and full release builds pass.
These are function timings, not FPS gains.

Only Matrix.cpp and MOSprite.cpp change in production. Fixture and reproduction instructions.

[Causeless]

Not sure we want to include the benchmarks stuff, but I don't see an issue with it for now