Add support for large `num_items` to `DeviceScan::*ByKey`

[elstehle]

Simply instantiating the scan_by_key kernel template with a 64-bit offset type turns out that it incurs some noticeable performance downside of -~35% (or 1.5-fold slowdown) . Hence, we want to also consider alternative approaches like the streaming approach or the bit-packed tile state.

Overview of performance different for various offset types compared to the currently used i32 offset type.

	Diff u32 vs i32 any num items	Diff u32 vs i32 2^28 num items	Diff i64 vs i32 any num items	Diff i64 vs i32 2^28 num items	Diff u64 vs i32 any num items	Diff u64 vs i32 2^28 num items
min	89.29%	89.29%	97.09%	99.94%	96.34%	100.00%
max	106.35%	102.38%	135.51%	131.50%	134.84%	131.56%
avg	100.06%	99.79%	107.28%	108.38%	107.31%	108.42%