Update OpenACC kernel timing and data regions

[marchdf]

Summary:

• Added guarded/optional NVTX ranges around the main OpenACC update phases for profiling
• Moved loop timing to include the full asynchronous kernel work by adding explicit acc wait points
• Used create(...) for unew, vnew, and pnew device data instead of copying uninitialized scratch arrays from host
• Added restrict qualifiers to the main field pointers to inform the compiler
• Removed misplaced parallel region incurred H2D and D2H memcopies of scalars
• Tuned the tiled OpenACC variant with separate TILE_I / TILE_J dimensions
• 1.05X to 1.2X speedup (non-tiled/tiled)

This is one of several PRs stemming from a discussion with @sjsprecious and @johnmauff.