[BUG] Extra `cuLibraryLoadData` calls when `LIBCUDF_HOST_DECOMPRESSION=AUTO`

[GregoryKimball]

Describe the bug
I was surprised to see extra calls to cuLibraryLoadData just before the first decompress_page_data range when LIBCUDF_HOST_DECOMPRESSION=AUTO is enabled. This happens even when CUDA_MODULE_LOADING=EAGER.

This load library region does not happen when LIBCUDF_HOST_DECOMPRESSION is unset or LIBCUDF_HOST_DECOMPRESSION=ON. In the PDS benchmark it adds perhaps 25 ms per query.

The library loading seems to be in sorting the blocks I guess SortPairsDescending.

Expected behavior
All the cuLibraryLoadData calls should be in the beginning when CUDA_MODULE_LOADING=EAGER

[GregoryKimball]

pds_q9_sf100.zip

[GregoryKimball]

It looks like the library loading for nvCOMP normally happens in the "dry run" execution that Polars benchmarks do https://github.com/pola-rs/polars-benchmark/blob/271ae57a2f6d1038821faf4e5f9e750548b9422b/queries/polars/utils.py#L66

However, when we are using AUTO, this dry run step no longer triggers nvCOMP library loading.