CUDA matrix reductions

See merge request exastencils/exastencils!55

- set cuda device according to knowledge flag

  - support pinned memory on the host side (active on default)
  - support managed memory
  - support zero copy
  - replace transfers with prefetch instructions in case of managed memory
  - eliminate transfers when zero copy is used

add cuda conditions inside of wrapping fragment loops for allocations (allows using fragmentIdx in condition)

VisIt coupling

See merge request exastencils/exastencils!46

Test commit: allow consecutive reduction compound assignments by introducing a kernel-local variable.

Add consecutive compound assignments to reduction variable within a kernel for CUDA_MatrixReduction test.

check for additional errors that are not reported by default in CUDA_CheckError

See merge request exastencils/exastencils!53

git guidelines

See merge request exastencils/exastencils!52

GitHub Page

See merge request exastencils/exastencils!51

Devel/fix matrix base dt

See merge request exastencils/exastencils!48

Distinguish between (non-)evaluable array/matrix accesses and setup device copy in the non-evaluable case.