Add a benchmark with unaligned inputs for `cub::DeviceTransform`

[bernhardmgruber]

The work on adding a Tile version of cub::DeviceTransform included a conversation that Tile wasn't performing great when the inputs were not aligned to a 16-byte boundary. The bulk copy kernel of cub::DeviceTransform contains special handling for such cases and it would be interesting to compare the performance of transforming unaligned inputs to aligned ones, as well as compare the bulk copy implementation with the tile kernel.