Add basic cuda/hip API interface

[fluidnumerics-joe]

API's exposed to end user (kaolin_array)

• Magic = for fortran allocatable to/from kaolin array
• copy_from (return fortran allocatable)
• copy_to

GPU Memory management

• gpumalloc
• gpumallochost
• gpufree
• gpumemcpy
• gpumemcpy

Data initializers

• zeros
• rand
• ones
• eye

For hand-written kernels and kernel wrappers, let's use chevron syntax for kernel launches <<<grid_size, block_size >>>

Tasks

• Look at stdlib pre-processor thingy to assess viability for selecting backend (cuda/hip) #2
• ??