This PR implements a basket of optimizations for the parquet reader to bring non-chunked reads close to par following the merge of the sub-rowgroup reader.

The primary culprit for the performance hit was that in the case where we perform no splits, we were making a full copy of all of the pages into the subpass struct (including a pinned memory allocation). This is unnecessary because we can just represent the pages in the subpass as a span that wraps the existing pages in the pass.

In addition, several hostdevice_vectors used for work that could be done entirely device-side were converted to rmm::device_uvector.

Finally, I converted a number of functions that were taking hostdevice_vectors to use spans instead and added some missing operators to the hostdevice_vector class itself.

This PR doesn't recover all the time (there is some new work that we have to do in all cases) but it takes out most of the sting. A sample of some of the benchmarks that were most notably affected:

                       Original Time      Sub-rowgroup-implementation       This PR
parquet_read_decode
Int, device buffer 0   29260860778        26373181343                       28121328587
Int, device buffer 1   30692134492        27474241282                       29495189226

parquet_read_chunks
Int, device buffer     33895028252        29986276949                       32293548191
Float, device buffer   57055985251        49640274260                       55795392897

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.