Fix JSONL row loss in sb-to-lines segment merge

Summary: The source-block-to-lines mapping is written by collect_source_block_to_lines_mapping via per-segment worker threads that each write a temp file, which a merger thread concatenates in order into the final JSONL. Two defects caused the merged output to silently drop the majority of emitted rows (measured at ~61% loss: the collector emitted 3,614,998 rows but only 1,417,120 reached the file).

1. Flush race: each worker signaled completion (finished_segments.insert(idx) + notify_all()) while its std::ofstream was still open — the stream is only flushed and closed when the lambda scope exits, after the signal. The merger could therefore read and std::filesystem::remove() a not-yet-flushed segment file, and the worker’s final flush would then land on an unlinked inode. Fixed by calling ofs.close() before signaling completion.

2. Failbit poisoning: the merge copy ofs << ifs.rdbuf() sets the output stream’s failbit whenever zero characters are inserted (an empty segment file, i.e. a segment whose methods produced no rows). Once the failbit is set, every subsequent ofs << ifs.rdbuf() is a silent no-op, dropping all later segments. Fixed by only performing the copy when the segment file is non-empty (ifs && ifs.peek() != eof); the file is still removed either way.

With both fixes the JSONL file now contains every emitted row.

Reviewed By: danjin250

Differential Revision: D113465381

fbshipit-source-id: 7f047d6557bc59ef5fb8e2285ffa0c76dca499fe