Support quoted multiline fields in parallel CSV parsing

[rahul-iyer]

This change adds a planned-range path for the parallel CSV reader so files with quoted newlines can
still use parallel parsing without splitting logical rows across workers.

The core change is a boundary-planning step that computes safe logical row ranges before worker parsing
begins. When MULTILINE_PARALLEL=true, the reader uses those planned ranges instead of naive fixed block
cuts. That lets workers parse independently while preserving correctness for multiline quoted fields.

What changed

• Added MULTILINE_PARALLEL as a CSV reader option.

• Added boundary planning for parallel CSV using CSVBoundaryScanner::planFixedChunkOverlap(...).

• Added csv_boundary_scanner.cpp to the CSV reader build target.

• Updated ParallelCSVReader to:

  • switch into planned-range mode when multiline parallel parsing is enabled
  • seek to planned range starts instead of fixed block boundaries
  • allow quoted newlines in planned-range mode
  • use the structural parser on planned ranges when supported

• Kept the existing fixed-block path unchanged for non-multiline parallel reads.

Benchmark Results

Median of 3 runs. Dataset results came from /private/tmp/lbug_dataset_benchmark_counts_after_simd.csv.
The synthetic comparison came from /private/tmp/lbug_mode_benchmark_results.csv. In the dataset runs,
outputs were identical across modes for every case listed below.

Dataset	File	Serial (s)	Parallel (s)	Parallel Speedup	Multiline Parallel (s)	Multiline Speedup	Output Match
amazon	dataset/snap/amazon0601/csv/amazon-edges.csv	0.196	0.047	4.21x	0.060	3.28x	Yes
twitter	dataset/snap/twitter/csv/twitter-edges.csv	0.175	0.044	3.97x	0.055	3.18x	Yes
comment	dataset/ldbc-sf01/Comment.csv	0.067	0.033	2.05x	0.038	1.76x	Yes
post	dataset/ldbc-sf01/Post.csv	0.063	0.031	1.99x	0.036	1.75x	Yes
facebook	dataset/snap/facebook/edges.csv	0.072	0.031	2.35x	0.037	1.93x	Yes
embeddings	dataset/embeddings/embeddings-960-1k.csv	0.077	0.059	1.31x	0.066	1.17x	Yes

Workload	Serial (s)	Parallel (s)	Parallel Status	Multiline Parallel (s)	Multiline Status	Notes
clean_256mb	1.046	0.152	Success	0.203	Success	Plain parallel is fastest on non-multiline input; multiline path still gives 5.16x over serial
multiline_256mb	0.783	0.089	Fails on quoted newlines	0.148	Success	Multiline parallel succeeds and is 5.30x faster than serial

[adsharma]

[edit - rebased and pushed ]
@rahul-iyer could you rebase your dataset module against origin/main? That should fix the test failure.

[adsharma]

@rahul-iyer only one failure on windows to fix

https://github.com/LadybugDB/ladybug/actions/runs/28725297352/job/85236841991

[adsharma]

Windows run with the proposed fix: https://github.com/LadybugDB/ladybug/actions/runs/28749519654