perf(import): convert JSON array rows lazily during import

[datlechin]

Follow-up to #1536 (post-merge review).

JSONImportPlugin.performImport built the entire array import into a [[String: PluginCellValue]] buffer (parseRows did extractRows(...).map(convertRow)) before inserting any row. For a large .json array that's a full duplicate of the row collection held alongside the parsed object graph.

This converts each row to PluginCellValue lazily at insert time instead: parseRows now returns the raw [[String: Any]] and the loop calls convertRow per row. Peak memory drops by roughly that extra buffer. No behavior change.

Note: the deeper cost (the whole file + JSONSerialization object graph in memory) is inherent to the array path; NDJSON/JSONL already streams via url.lines. A streaming array parser would be a separate, larger change.

Tests unchanged (they exercise extractRows/convertRow/parseRow directly, none of which changed signature). Local full-suite validation was blocked by an unrelated DerivedData/SPM checkout issue on my machine; relying on CI's clean build here.