This commit addresses the performance issue reported in GitHub issue opencv#25928 where
cv2.norm(img1, img2) became significantly slower in versions newer than 4.2.0.34.

The root cause was identified as the change that moved IPP optimizations to the
Hardware Abstraction Layer (HAL) in PR opencv#27128. This architectural change
improved code structure but introduced overhead for this specific case.

Changes include:
- Add direct SIMD-optimized implementations for L1, L2, and INF norms for float matrices
- Add performance tests to verify the improvement
- Add documentation explaining the fix

With this change, cv2.norm(img1, img2) should be faster than the workaround
of using cv2.norm(img1 - img2), which is important for image comparison operations.

Fixes: opencv#25928