Initial DeviceScan deferred num_items support only for look-ahead and non-deterministic

Add initial support for device-resident num_items in cub::DeviceScan, limited to the Warpspeed look-ahead path in non-deterministic mode. The kernel will read num_items from the device-resident source on entry, derive num_tiles and its per-tile work range on device, and perform per-tile prefix propagation through the existing look-ahead global tile-state array. For temp-storage for the tile-state buffer we can use a caller-provided upper bound on num_items (via cuda::args::deferred bounds) or something like INT_MAX.

This issue can be closed once the look-ahead non-deterministic works locally with minimal testing.