enhance: pin sealed read-snapshot view reads through frozen column - #53913

[P2] Avoid rebuilding skip-metrics resources for every data read

The new data helpers consume only .first, but SegmentReadSnapshot::GetDataScanResources also constructs a FieldSkipMetricsView: its implementation copies the column shared_ptr twice and calls GetSkipMetricsList() before the unused .second is destroyed. Thus every chunk/view read still performs two reference-count increment/decrement cycles, and lazy storage-v2 columns also pay an acquire load and metrics lookup, so the hot paths do not get the zero-refcount access this change is intended to provide. Add a column-only snapshot accessor that returns a borrowed pointer/reference (the snapshot already owns the published state) and use it throughout this helper family, SegmentChunkReader, and group-by.

[P2] Carry the snapshot into strict phase-2 membership scans

The snapshot is passed only to the phase-1 data getter here. In the PerGroup branch, BuildGroupOffsets still creates a fresh SegmentChunkReader without calling SetSnapshot and reads chunk metadata directly from the segment, so its full raw-field scan continues to recapture published state for each chunk-size/data transition. Since this branch can scan total_data_cnt_ rows, it leaves the most expensive strict group-by path outside the optimization. Carry the snapshot in StrictGroupPhase2Context, pass it to BuildGroupOffsets, and bind both the reader and metadata lookups to that snapshot.

[P3] Place the common snapshot argument before JSON-only options

Appending snapshot after the three defaulted JSON-specific parameters forces every scalar branch to pass the same std::nullopt, std::nullopt, and false placeholders, expanding simple one-line cases into most of the new switch plumbing. This helper is file-local, so move snapshot before those optional parameters or capture it in a local factory; the scalar cases can then remain direct calls without dummy arguments, and the JSON cases keep supplying only their real options.

[P2] Keep the fallback column owner alive until pinning

When snapshot_ is null, GetChunkedColumn() returns a shared_ptr, but calling .get() here keeps only its raw pointer and destroys the temporary owner at the semicolon. A concurrent sealed-state publication can then retire that column before column->GetChunk() runs, causing a use-after-free in the supported non-pinned fallback. The previous implementation kept the shared_ptr alive through pinning. Retain an owning shared_ptr in the fallback branch until GetChunk() returns, and use a borrowed pointer only for the snapshot-backed branch.

[P3] Remove the duplicate sealed-view wrapper layer

GetBatchViews and GetViewsByOffsets only forward to the existing snapshot-aware GetChunkView and GetChunkViewsByOffsets helpers, or to segment aliases with the same behavior. All three call sites already execute only for sealed segments, so this additional API layer does not enforce a reachable contract and duplicates logic that must remain synchronized. Call the existing helpers directly at those sites, passing the range where needed, and remove both wrappers.

[P3] Reuse the reader for snapshot-aware metadata

The new membership path selects between snapshot and segment separately for chunk count, row prefixes, and every chunk size, then immediately creates a SegmentChunkReader and binds that same snapshot. Move reader construction and binding before these checks and use NumChunkData, NumRowsUntilChunk, and ChunkSize; this removes duplicate generation-selection logic and prevents the metadata scan from drifting from the accessor's source.

[P3] Avoid caching a one-use snapshot alias

snapshot_ duplicates state already owned by query_context_ and is read only once in GetOutput, which can execute its output-producing path only once. Obtain a local shared snapshot from query_context_ at that use site and pass its pointer to SearchGroupBy; then remove this raw-pointer member and its constructor assignment so ownership and lifetime remain visible at the call.

[P3] Centralize the duplicated snapshot column adapters

The new GetChunkView and GetChunkViewsByOffsets blocks duplicate the type dispatch and JSON materialization already implemented by SegmentInternalInterface::chunk_view and chunk_views_by_offsets, while the same SpanBase conversion is repeated again in SegmentChunkReader and SealedDataGetter. This leaves multiple supported-type matrices that must stay synchronized and makes a new view type easy to add to one path but not the others. Extract the column-backed typed adapters once next to ChunkedColumnInterface, then let these callers only select the snapshot column or the existing segment fallback.