Optimize parquet reader mukernel block scans, non-nullable code

[pmattione-nvidia]

When doing block scans in the parquet reader we also need to know the per-warp results of the scan. Because cub doesn't return those, we then do an additional ballot that is unnecessary. This introduces code that does a block scan manually, saving the intermediate results. Also by doing an exclusive-scan instead of an inclusive-scan, we don't need the extra "- 1's" that were everywhere.

For (non-list) nested columns that aren't nullable, we don't need to loop over the whole nesting depth; only the last level of nesting is used. After removing this loop, the non-nullable code for nested and flat hierarchies is identical, so they're extracted and consolidated into a new function.

By having each thread keep track of the current valid_count (and not saving-to or reading-from the nesting_info until the end), we don't need to synchronize the block threads as frequently, so these extra syncs are removed.

Checklist

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