perf: add low-overhead row decode APIs for Node.js

[Nebu1eto]

Summary

• Add column name cache to existing decodeRowsBuffer() eliminating redundant string computations
• Add getRowsRaw() API returning typed arrays instead of JS objects (zero per-cell object allocation)
• Add getRowsIterator() generator for lazy row-at-a-time processing without full materialization

Performance Impact

• Column name cache: Reduces O(rows * cols) string computations to O(cols) unique lookups
• getRowsRaw(): For 50k x 20 sheet, creates 7 typed arrays instead of 50,000 JsRowData + 1,000,000 JsRowCell objects
• getRowsIterator(): Avoids holding all rows in memory simultaneously
• Target: Node Read Multi-Sheet >= 10% improvement

New APIs

// Typed-array based (zero object allocation)
getRowsRaw(sheet: string): RawRowsResult

// Generator-based (lazy, streaming)
*getRowsIterator(sheet: string): Generator<JsRowData>

Changes

• packages/sheetkit/buffer-codec.ts: Column cache, decodeRowsRawBuffer(), decodeRowsIterator(), shared parseBuffer()
• packages/sheetkit/index.ts: getRowsRaw(), getRowsIterator() methods + RawRowsResult export
• packages/sheetkit/__test__/index.spec.ts: 22 new tests across 4 describe blocks

Test plan

• 22 new tests (getRowsRaw, decodeRowsRawBuffer, getRowsIterator, decodeRowsIterator)
• All 486 existing tests pass (includes 115 Node.js tests + Rust tests)
• Cross-validation: getRowsRaw results match getRows for all cell types
• No Rust changes required - pure JS/TS optimization

🤖 Generated with Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

[Nebu1eto]

Review finding (performance regression): column-name cache prewarm penalizes sparse/wide sheets.

decodeRowsBuffer() and decodeRowsIterator() now precompute all column names up to maxCol before decoding:

• packages/sheetkit/buffer-codec.ts (cachedColumnName(maxCol) in decodeRowsBuffer)
• packages/sheetkit/buffer-codec.ts (cachedColumnName(maxCol) in decodeRowsIterator)

For sparse buffers, this changes complexity from roughly O(non-empty-cells) column-name work to O(colCount). In this format, colCount is the full bounding range width (from minCol to maxCol), not the number of populated columns.

I reproduced this on a sparse wide sheet (A1, XFD1):

• main decodeRowsBuffer first call: ~0.435 ms
• this PR decodeRowsBuffer first call: ~1.599 ms
• follow-up call is faster due to warmed global cache, but first-call latency regresses by ~3.7x.

This also weakens iterator behavior: decodeRowsIterator() now does non-trivial up-front work before first next(), which is opposite to expected lazy/streaming ergonomics.

Suggested fix:

• Only prewarm for dense layout (or cap prewarm to a small threshold).
• For sparse layout, resolve column names lazily per encountered cell using the cache-backed function.