Correct handling of solve_type

[rbabich]

As originally conceived, a "DIRECT" solve_type is supposed to result in the matrix M being inverted directly, while a "NORMEQ" solve inverts M^† M. From the NEWS file:

• Added 'solve_type' to QudaInvertParam. This determines how the linear system is solved, in contrast to solution_type which determines what system is being solved. When using the CG inverter, solve_type should generally be set to 'QUDA_NORMEQ_PC_SOLVE', which will solve the even/odd-preconditioned normal equations via CGNR. (The full solution will be reconstructed if necessary based on solution_type.) For BiCGstab, 'QUDA_DIRECT_PC_SOLVE' is generally best. These choices correspond to what was done by default in earlier versions of QUDA.

At the moment, this doesn't work as advertised; CG always does a "NORMEQ" solve, irrespective of the value of solve_type.

It also wouldn't hurt to choose a more sensible naming convention (at the very least replacing "NORMEQ" by "NORMOP"), as advocated by Balint.

[gshi]

Ron, Could you add a complete description of which type solves which equation using which method somewhere in the code or document? I find myself always confused, even after I thought I had everything understood. Thanks

[maddyscientist]

Moved to 0.4.1 as this can be worked around for now.