Speed up .sng header parsing (~3.3×) and add readSongIni for metadata-only scans

[elicwhite]

Summary

A series of focused, independently-reviewable perf improvements to the .sng header parse path, plus one small additive API (readSongIni) for the common "I only want the song.ini data" use case.

Each commit is scoped to one change, with profiling rationale and before/after benchmarks in the commit body. Nothing breaks existing behavior — SngStream's public contract is unchanged.

Cumulative impact (204 .sng files, concurrency=32, node 24.11):

Scenario	Before	After	Delta
Header parse via SngStream (cold cache)	11.51 ms / 2,540 files/s	3.32 ms / 8,303 files/s	−71% latency, +227% throughput
Metadata-only scan (warm cache, readSongIni)	2.23 ms / 13,510 files/s (via event API)	1.63 ms / 18,820 files/s	−27% latency, +39% throughput

Commits, in review order

1. Replace chunked header buffer with single growing buffer — the biggest win. The old getHeaderBuffer scanned every byte of every chunk on each of three length probes, and mergeUint8Arrays used an O(n²) .slice(0, i).reduce(...) inside .forEach. Replaced with a single geometrically-growing Uint8Array. −48% latency, +88% throughput.
2. Hoist binary-parser Parser definitions to module scope — parseSngHeader was constructing three binary-parser.Parser instances on every call, throwing away binary-parser's internal compiled-function cache. Moved to module scope. −8% latency, +9% throughput.
3. Skip generateIniFileText when generateSongIni is disabled — generateIniFileText was building + encoding ini text on every header parse even when the result was discarded. Moved inside the generateSongIni: true branch. −10% latency, +9% throughput.
4. Defer leftover-chunk unmask from start() to pull() — when the header's final read() overshoots into the first file's bytes, readFile's ReadableStream.start was synchronously XOR-unmasking those leftover bytes for ~40–60 KB per file, before a metadata-only consumer's cancel() could land. Deferred to the first pull() so cancel-before-pull skips the XOR entirely. −33% latency, +47% throughput for metadata-only scanners; identical behavior for full-extraction consumers.
5. Add readSongIni() for metadata-only scanners — a single new top-level export. Reads just enough of a stream to parse the header, cancels the source, returns { [key: string]: string } (same shape as SngHeader.metadata). Skips EventEmitter, file ReadableStream construction, chunk-unmasker allocation, and readFile dispatch — all unnecessary when the caller only wants the header. −27% latency, +39% throughput vs. using SngStream + cancel for the same workload.
6. Add changeset for header-parse perf improvements — bumps the package minor (new public API).

Testing

• tsc --noEmit clean
• tsup build succeeds
• End-to-end extraction of a multi-file .sng with generateSongIni: true produces byte-identical output at each commit (MIDI/JPEG/Opus magic bytes intact, sizes match, generated song.ini well-formed)
• readSongIni returns the same metadata object shape as SngHeader.metadata
• CPU profiling was used at each step to verify the expected hotspot was eliminated