Parameterize cholinv

[huttered40]

A few design decisions deserve to be parameterized instead of forced, most notably the use of gemm vs. trmm and the corresponding serializations.

Any policy classes must have default types.

Note that validate classes and all code that references choleskyqr must also follow these rules.

[huttered40]

I found another place in which we are not exploiting the MatrixType template parameters and could use another parameterization (perhaps in place of the dummy one I already set up): at the end, where we perform the last matrix products, we are not exploiting the triangular structure of MatrixRI, and then we serialize from square to square. What is that about?

[huttered40]

See #21 for more information

[huttered40]

Add a template parameter for OverlapGather, with the default being no overlap.

[huttered40]

• Add template parameter for GEMM vs. TRMM
• Add template parameter for packed vs. non-packed in tradeoff of less communication and more serialization.

[huttered40]

I removed the OverlapGather policy, as I was seeing no benefit (actually made things worse) and also because it was not correct for num_chunks>1 (I don't want to debug it).

We do not need a GEMM vs. TRMM policy, as I think its safe to assume that TRMM's performance is nearly that of TRMM, and when taken into account the less flops for TRMM, there is no tradeoff benefit in choosing GEMM.