Preconditioner cleanup

[weinbe2]

This narrow PR cleans up various bits and bops related to how preconditioners are created and handled in QUDA, plus other various helpful utilities:

• Creates an abstract getInverterType() function within the base Solver class, forcing each derived class to specify its QudaInverterType enum.
• Created a single Solver function fillInnertParam/extractInnerParam to handle copying preconditioner properties in/out, which removed redundant code in GCR and PCG
• Unified creating preconditioners, plus wrapping them in std::shared_ptr to simplify cleanup
• ...which requires some trickery, a shared pointer with a custom no-op deallocator, to handle MG preconditioners
• Threaded in support for an MG preconditioner for PCG as a proof-of-concept (even though there's no formal operator with a Hermitian coarse op in QUDA)

Outstanding bits:

• [x] clang-format
• [x] ctest

[weinbe2]

This now passes ctest. @maddyscientist , assuming this now passes Jenkins, can you give this a look (and I can subsequently clang-format it)?

[maddyscientist]

This looks fairly inoffensive. Regarding ctest passing: does this pass on multiple GPUs? Checking, since this PR touches things with Schwarz solvers, just checking that the Schwarz-preconditioned tests are working ok (Mobius PCG and Wilson GCR).....

[weinbe2]

This looks fairly inoffensive. Regarding ctest passing: does this pass on multiple GPUs? Checking, since this PR touches things with Schwarz solvers, just checking that the Schwarz-preconditioned tests are working ok (Mobius PCG and Wilson GCR).....

Yes, my ctest experiments were all with -DQUDA_MPI=ON builds running on 8xGPUs, 1x2x2x2 grid partitioning.