[BUG] use device_uvector instead of device_vector when possible

[vuule]

The new device_uvector does not initialize the allocated memory so it's much faster when the initial content is not relevant.
There are many places where device_vector can be replaced with device_uvector to improve the performance. Example:

rmm::device_vector<size_type> left_indices(left.num_rows());
thrust::sequence(
  rmm::exec_policy(stream)->on(stream), left_indices.begin(), left_indices.end(), 0);

[jrhemstad]

Furthermore, device_uvector has a stream argument in its constructor which allows correctly stream ordering the allocation. This is currently not possible with device_vector and can lead to significant issues when using non-default streams (see #2631).

[jrhemstad]

Didn't see this when I wrote #7287. Going to close this as a duplicate since I provided a little more context in that issue.