Optimize list_contains with prepared constant sets - #10052

Reverts faad45c. That kernel decompressed a Zstd needle once and probed
a hash set of a constant list, because the generic path compared every
list element against the needle and decompressed it each time. Upstream
vortex-data#10052 now reduces a constant list to a PreparedSet, which canonicalizes
the needle once and probes its own hash set, so the kernel only adds a
decompression step before the same lookup.

encodings/zstd/benches/list_contains.rs (8k Zstd strings, median),
list_contains over the Zstd values:

  list length   with kernel   without
  4             147.5 µs      134.7 µs
  256           160.6 µs      146.9 µs
  2,048         195.8 µs      183.3 µs
  24,999        792.3 µs      725.1 µs

Before vortex-data#10052 the same call took 29 ms at 2,048 elements when the values
were decompressed first, and 252 ms when they were not.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Reverts faad45c. That kernel decompressed a Zstd needle once and probed
a hash set of a constant list, because the generic path compared every
list element against the needle and decompressed it each time. Upstream
vortex-data#10052 now reduces a constant list to a PreparedSet, which canonicalizes
the needle once and probes its own hash set, so the kernel only adds a
decompression step before the same lookup.

encodings/zstd/benches/list_contains.rs (8k Zstd strings, median),
list_contains over the Zstd values:

  list length   with kernel   without
  4             147.5 µs      134.7 µs
  256           160.6 µs      146.9 µs
  2,048         195.8 µs      183.3 µs
  24,999        792.3 µs      725.1 µs

Before vortex-data#10052 the same call took 29 ms at 2,048 elements when the values
were decompressed first, and 252 ms when they were not.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Since upstream vortex-data#10052, optimizing `list_contains(literal_list, needle)`
replaces the literal list with a `PreparedSet`. `list_contains_falsify`
read the elements only from a literal, so once the scan had optimized its
filter, stats pruning gave up on every IN list: no whole-file pruning from
footer statistics and no zone-map pruning. On SmithDB dev, a 50-trace-ID
lookup read 12.8-15.9 MB of zone-map and Bloom segments where release-13
reads ~50 KB.

Read the elements from the prepared set as well.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Reverts faad45c. That kernel decompressed a Zstd needle once and probed
a hash set of a constant list, because the generic path compared every
list element against the needle and decompressed it each time. Upstream
vortex-data#10052 now reduces a constant list to a PreparedSet, which canonicalizes
the needle once and probes its own hash set, so the kernel only adds a
decompression step before the same lookup.

encodings/zstd/benches/list_contains.rs (8k Zstd strings, median),
list_contains over the Zstd values:

  list length   with kernel   without
  4             147.5 µs      134.7 µs
  256           160.6 µs      146.9 µs
  2,048         195.8 µs      183.3 µs
  24,999        792.3 µs      725.1 µs

Before vortex-data#10052 the same call took 29 ms at 2,048 elements when the values
were decompressed first, and 252 ms when they were not.

Co-Authored-By: Claude Opus 5.5 <[email protected]>