Add GPU construction for 32-bit Roaring bitmaps from raw indices

[sleeepyjack]

• add from_indices, from_sorted_indices, and from_sorted_unique_indices factories for constructing portable 32-bit Roaring bitmaps from device-accessible ranges of raw indices
• use CUB for sorting, deduplication, container discovery, and workspace reuse, with custom array/bitset serialization kernels and one synchronization for exact output sizing
• add from_serialized for consistency with the other factories as an alternative to the existing ctor which will be removed in the future, fix empty bitmap parsing and cached references after moves, and add tests, examples, and NVBench coverage

Deferred features:

• no run container support since it is up-to 24x slower during construction and is optional
• additional optimizations for small bitmaps aka run the whole pipeline in a single CTA
• no batch build yet (TBD)