Organize parquet reader mukernel non-nullable code, introduce manual block scans

[pmattione-nvidia]

This is a collection of a few small optimizations and tweaks for the parquet reader fixed-width mukernels (flat & nested, lists not implemented yet). The benchmark changes are negligible, this is mainly cleanup and code in preparation for the upcoming list mukernel.

1. If not reading the whole page (chunked reads) exit sooner

2. 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.

3. 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.

4. 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 warp-wide ballot that is unnecessary. This introduces code that does a block scan manually, saving the intermediate results. However using this code in the flat & nested kernels uses 8 more registers, so it isn't used yet.

5. By doing an exclusive-scan instead of an inclusive-scan, we don't need the extra "- 1's" that were everywhere.

Checklist

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

[nvdbaranec]

First pass. Will absorb and come back for another.

[vuule]

Nice set of changes!