Maximal vectorization

[maddyscientist]

This PR is a significant cleanup and reworking of the native QUDA accessors.

• Gone are explicit FLOAT2, FLOAT4, FLOAT8 data orderings, we now just have NATIVE ordering
• The data ordering is now set for all fields by CMake parameters
  • QUDA_ORDER_DOUBLE, QUDA_ORDER_SINGLE, QUDA_ORDER_HALF, QUDA_ORDER_QUARTER
  • The values of these correspond to the inner vector length desired, e.g., 4 would be a FLOAT4 accessor
• For fields, whose degrees of freedom are not a multiple of the vector length we deal with the remainder explicitly
  • E.g., for a SU(3) field with 18 real numbers, and a FLOAT4 accessor, we would have 4x FLOAT4 ld/st instructions and a 1x FLOAT2 remainder.
• Vector lengths 2, 4, 8, and 16 are supported (up to 256-bit in total length)

The motivation for this work is to increase the use of vectorized load and stores, to improve performance on more recent GPUs.