feat: multi-RG output at metric_name boundaries (PR-6c)

[g-talbot]

Summary

• Slices each merge output into row groups at metric_name transitions, greedily packing adjacent runs until adding the next would exceed MergeConfig::writer_config.row_group_size.
• Outputs that fit in one RG remain single-RG (vacuously sort-prefix-aligned). Outputs that exceed the cap roll into multiple RGs, every boundary aligned to a metric_name transition.
• A single metric_name run that on its own exceeds the cap stays as one RG — splitting it would break sort-prefix alignment, which the slicer prefers to preserve over the row-count budget.

Implementation

• pack_into_metric_aligned_rgs(batch, max_rg_rows) finds metric_name run boundaries (handles both Utf8 and Dictionary(_, Utf8) since post-align_inputs_to_union_schema is Utf8 and optimize_output_batch may dict-encode for low-cardinality strings) and runs a greedy packer.
• column_major_write issues one start_row_group / column writes / finish cycle per RG range; columns are zero-copy Array::slice views into the parent sorted batch.
• If the output schema is missing metric_name (defensive — production schemas always include it), falls back to single-RG with a debug_assert!(false, ...).

What's not in this PR

• qh.rg_partition_prefix_len metadata propagation. The marker constant lives on PR-1's branch (gtt/parquet-page-stats-and-marker, not yet merged); PR-6c produces metric-aligned output structurally and PR-7 (wiring) will add the metadata claim once PR-1 lands. Until then PR-6c outputs look prefix=0/absent to readers — they'll route through the legacy adapter unnecessarily but correctly.

Stack

main
└── gtt/parquet-streaming-base
    └── gtt/column-page-stream-trait (PR-5a #6406)
        ├── gtt/legacy-input-adapter (PR-5  #6408)
        └── gtt/parquet-page-decoder  (PR-6a #6407)
            └── gtt/streaming-merge-engine-merger (PR-6b #6409)
                └── gtt/streaming-merge-engine-multi-rg ← PR-6c (this PR)

Test plan

• test_multi_rg_at_metric_name_boundaries — three metric_name runs ≤ row_group_size produces 3 RGs each containing one metric_name
• test_metric_name_runs_packed_into_single_rg_when_under_threshold — three small runs total ≤ threshold pack into 1 RG
• test_metric_name_runs_pack_then_roll — pack until the next run pushes past threshold, then roll
• test_single_run_larger_than_threshold_stays_one_rg — 250 rows, threshold 50 → still 1 RG (alignment preserved)
• test_multi_rg_total_rows_preserved — MC-1 across 4 metric_names with mixed sizes
• test_multi_rg_output_drainable_by_stream_decoder — round-trip output → StreamDecoder yields one batch per RG

CI gates locally green: clippy --workspace --all-features --tests with -Dwarnings, nightly fmt --check, cargo doc --no-deps, cargo machete, license headers, log format, typos. 417/417 crate tests pass (16 streaming tests, +6 new from PR-6c).

🤖 Generated with Claude Code