Propagate `@inbounds` across kernel boundary to make array operations safer

[maleadt]

sortperm! takes indices as inputs, and preforms @inbounds accesses that can crash the GPU:

julia> CUDA.@sync sortperm!(cu([1,10000000]), CUDA.rand(2); initialized=true);
ERROR: CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)

Running this with --check-bounds=yes reveals that the bitonic sort implementation takes these input indices and blindly uses them to index shared memory. That's bad.

Normally we solve this by marking the function @propagate_inbounds instead of @inbounds, and expect the user to vouch for the inputs. However, @inbounds doesn't propagate across kernel launch boundaries.

[vchuravy]

We could inject Expr(:boundscheck) into the @cuda and then propagate that to a compiler argument.