MMA-optimized batched MRHS MG

[maddyscientist]

This PR focuses on improving the use of MMA with batched multigrid, resulting in the optimal MRHS multigrid performance:

• Make the mma-ordered version of the null space vectors persistent to avoid needing to reorder on the fly
• Full coarse-grid correction can now be done in MMA order, minimizing the reordering overhead
  • This is enabled on a per-level basis using QudaMultigridParam::collapse_mrhs (--mg-collapse-mrhs from the command line)
  • Doing so collapse the MRHS solve to a single solve on the coarse level in question
  • Enabling this requires that both the transfer operation into that grid and the dslash operator for that grid are deployed using MMA
• Improve memory utilization: if pre- and post- smoother are identical, they can alias, avoiding needing to be allocated separately
• Various boiler-plate improvements to facilitate the above, e.g.,
  • input and/or output for transfer operators can be collapsed
  • new method vector_ref<ColorSpinorField>::size_actual() to allow us to query the number of RHS in a collapsed system
• Improve usability of FieldTmp
  • Temporaries can now be references, which helps with composability
• BlockTranspose input field set can now be in half precision