CUDA kernels that are implemented but not optimal

[jpivarski]

This is primarily for record-keeping, so that we don't forget about CUDA kernels that should be revisited someday. To be in this list, a kernel must be implemented correctly (in main or an impending PR), but have some reason to be rewritten. The list is to help us stick to the policy that existence is the first priority and optimization is second, without the temptation to go down a rabbit-hole of optimizing every kernel before moving on to the next one.

• awkward_ListArray_min_range is a reduce-to-scalar algorithm, but it's implemented with atomicMin instead of tree-reduction.

Variable-length inner loop:

• awkward_IndexedArray_ranges_next_64
• awkward_IndexedArray_ranges_carry_next_64
• awkward_ListArray_getitem_jagged_numvalid
• awkward_ListArray_getitem_next_range_spreadadvanced
• awkward_ListArray_broadcast_tooffsets
• awkward_ListArray_localindex
• awkward_ListOffsetArray_drop_none_indexes
• awkward_ListOffsetArray_reduce_local_nextparents_64
• awkward_ListArray_rpad_axis1
• awkward_ListOffsetArray_rpad_axis1
• awkward_ListArray_combinations_length
• awkward_NumpyArray_pad_zero_to_length
• awkward_NumpyArray_rearrange_shifted
• awkward_UnionArray_flatten_combine
• awkward_UnionArray_nestedfill_tags_index