Track CuArray stream usage

[maleadt]

We currently require that tasks where CuArrays are written to, synchronize the GPU before sending that data off to other tasks. This is fragile, difficult to explain, and a source of common errors. We should consider tracking which tasks or streams a CuArray is used in, in order to automatically synchronize on use.

[maleadt]

cc @jpsamaroo; this would also help Dagger.

One potentially cheap way to do this is by keeping track of a single stream in each array, and when deriving a pointer (which dominates every use of a CuArray) check whether the current stream matches the array's. If not, synchronize the contained stream and replace it by the current one.