Fix Grace-specific test failures:

1. Add stream syncs at the places where host-allocated memory may be destructed while still being accessed by GPU to avoid relying on stream-ordered semantics of memory allocations.
2. A bug in tests: CAGRA index produced by cagra::build is not guaranteed to be owning. The tests assumed otherwise; when the host dataset is passed, and it's accessible on the device (it's the case with Grace), the created index ended up non-owning. The lifetime of the host dataset in the tests was smaller than of the index, which led to invalid host accesses from the device.
3. A bug in dataset_deserialize.hpp: deserialize_strided function constructed a non-owning strided dataset, because the host data was accessible by the GPU. The current fix is to add a move-semantics overload of make_strided_dataset that always owns the passed data (either via moving the mdarray or by copying the data).