Epic: Time series data support #10171

Summary: support time series data, i.e. tables sorted by a timestamp column with high-volume appends, range queries over time, and time-bucketed aggregations.

Sortedness

• Record a per-file sort key in the footer and fill is_sorted / is_strict_sorted in file stats.
• Binary search or zone seeking on the sort key for time-range predicates.
• Reverse scan for newest-first queries.
• Infer DataFusion output_ordering from the file sort key instead of requiring WITH ORDER, and support ordering on the v2 table.
• Let the compressor use IsSorted to pick Sequence or Delta without trial encodes.

Aggregate pushdown

Filter and limit pushdown

• Allow filter and limit in the same scan and short-circuit once the limit is met.
• Push limit from DataFusion when a filter is present.
• Push TopK and join dynamic filters into zone pruning, matching DuckDB.
• Push Interval and Duration literals and now() - interval style predicates from DataFusion.

Better timestamp and other types

• Add Duration and Interval dtypes and fix the Arrow mappings that currently panic.
• Cast between timestamp units.
• Coerce time units in comparison and pruning instead of erroring.
• Finish DateTimeParts comparison kernels so sub-day range filters do not fall back to a full decode.

New scalar functions

• date_trunc, date_bin, date_part / extract, timestamp plus or minus interval.
• Push these down from DataFusion as vortex-expr.

Pruning extensions

• Falsifiers for functions of the time column, starting with date_trunc / date_bin.
• Multi-column predicates such as interval overlap.
• Wire the existing Bloom filter aggregate into the default writer for tag columns.

Adding delta

• Enable Delta in an edition and fix its cast rule.
• Delta-of-delta for near-regular timestamps.
• Re-apply Delta to monotone children like RLE indices and string offsets.

Benchmarks

• Sorted ClickBench: hits sorted by EventTime, same queries.
• TSBS DevOps or IoT style workload: last-point, time-range, group-by-bucket and newest-N queries.
• Compression-ratio suite on regular and jittered timestamps, counters and sensor doubles against Parquet, Gorilla and Pco.
• Ingest benchmark: streaming append throughput and time to first readable row.