feat: add sorted_series column for DataFusion streaming aggregation

[g-talbot]

Summary

• Compute a composite, lexicographically sortable sorted_series binary column at Parquet write time using storekey order-preserving encoding
• For each row, encodes non-null sort schema tag columns as (ordinal: u8, value: str) pairs, then appends timeseries_id (i64) as final discriminator
• Identical timeseries always produce identical byte keys regardless of timestamp or value, enabling DataFusion's streaming AggregateExec and BoundedWindowAggExec with O(1) memory instead of O(N) hash tables
• Column is placed after sort columns in physical layout (Phase 1b in reorder_columns) for optimal streaming read
• Fixes create_nullable_dict_array bug: dictionary keys now correctly index into unique values (was using original array index, causing panics for mixed null/non-null inputs)

Stacked on top of #6287 (column ordering) and timeseries_id work.

Design

Based on the Sorted Series Column design doc:

Key structure for sort schema [metric_name(0), service(1), ..., host(5)]:

┌──────────┬────────────────┬──────────┬──────────────┬─────────────────┐
│ ordinal 0│ "cpu.usage"    │ ordinal 1│ "api"        │ timeseries_id   │
│ (u8)     │ (storekey str) │ (u8)     │ (storekey)   │ (storekey i64)  │
└──────────┴────────────────┴──────────┴──────────────┴─────────────────┘

Null columns are skipped. The ordinal prefix prevents cross-column byte collisions for sparse schemas.

Test plan

• 19 tests (identity, discrimination, sort-order, null handling, stability, Parquet round-trip, structural ordinal verification, 2 proptests)
• All 195 quickwit-parquet-engine tests pass
• Clippy clean, formatted, no unused deps
• License headers pass
• Docs compile

🤖 Generated with Claude Code