As discussed in #882:

The LuDecomposition and CholeskyDecomposition classes expose properties with a simple Lazy evaluation pattern on the factored matrices. The QrDecomposition does not do this. Any unsuspecting user attempting to iterate through, say, the UpperTriangleFactor directly is in for a nasty shock as the matrix could potentially be rebuilt thousands of times.

Propose lazy cache on QR properties (or method call if caching undesirable). On phone but can post code snippet later if this is not clear. SVD has a few issues as well.

Fixed a few others and added checks in the unit test, Assert.AreSame() to check instance equality.

Thanks,
Alex

This closes #882