RHS for SVD and QR approaches are m x m, but can be reduced

[kuberry]

If SVD or QR are used, the RHS could really just be a m x 1 because it is a column scaling for U^T.
Because this matrix also stores the eventual polynomial coefficient matrix, it must be at least m x n.
Currently this is allocated to m x m which is unnecessarily large.

If LU is used, there is no issue because P rather than RHS is used for the actual system right hand side solve, which already is m x n.

This should be addressed as part of the conversion from LAPACK/CUBLAS implementation to batched QR with pivoting conversion in KokkosKernels.

[kuberry]

The above is currently completed in the k2_qr_pivot and qr_pivot branch.

Remaining things to do:

• Reduce number of items using scratch space and requiring a copy
• Check for serial case, and run it differently (no teams)
• Separate work vector in UTV as level 0, and in SolveUTV as level 1. For LU keep work vector for SolveUTV as _N x _NRHS, while for SVD or QR set it to size 0 and pass in level 0 work vector that is size _M and copy of contents of B prior to using B to store solution. Try to use B matrix (n x m) to store (U'*B)
• Clean up print statements in QR debug
  - [ ] Consider changing THREADS= to TEAM_THREADS= and VECTORLANES= to TEAM_VECTOR_LANES=