Fixes discard_memory compilation failure for pre-Volta

[elstehle]

Description

Closes issue #624 and unblocks PR #619

Including <cuda/annotated_ptr> includes <cuda/barrier>, which in turn includs "std/barrier". "std/barrier" in turn has the following guard in place:

#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 700
#  error "CUDA synchronization primitives are only supported for sm_70 and up."
#endif

This prevents users from using functionality like cuda::discard_memory in <cuda/annotated_ptr> that does not rely on barrier.

This PR:

• moves cuda::discard_memory to its own header
• fixes aligned pointer computation in cuda::discard_memory and the loop for discarding memory
• fixes initialization order in __access_property that the compiler complained about

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.