[batch-matrix-multiply] Guard residual stride rounding - #11402

Summary

• Add checked rounding for residual k/n stride, weights stride, alignment, input stride, and workspace-offset calculations.
• Reject overflow in batch tiling and final GEMM stride setup.
• Add n-stride and k-stride regression tests.

This complements #11400, which propagates qpint8 packed-size sentinels through the same operator.

Tests

• cmake --build build --target batch-matrix-multiply-nc-test -j2
• ctest --test-dir build -R '^batch-matrix-multiply-nc-test$' --output-on-failure