v5: split into PbfReader/PbfWriter; 25% faster

[mourner]

Breaking change: the single Pbf class is split into PbfReader and PbfWriter, allowing readers and writers to be tree-shaken independently in bundled apps.

Removed: the legacy readVarint64 method (replaced by readVarint(true) for signed 64-bit reads).

Performance: several targeted optimizations to both read and write hot paths, validated against bench/bench-tiles.js:

• makeRoomForExtraLength uses Uint8Array.copyWithin instead of a manual byte-shift loop (biggest write win).
• writeVarint gets a single-byte fast path, skipping the multi-byte branches and full realloc(4) for the common case (tags, small ints).
• readVarint gets a matching single-byte fast path.
• writeBytes uses typedArray.set instead of a manual loop.
  README updated with fresh medians (3 runs of bench/bench-tiles.js, Node v26).
• ~18% faster decode for the Mapbox vector-tile workload (~236 → ~200 ms, 159 → 187 MB/s) by rewriting reader codegen from a callback-style pbf.readFields(cb, obj, end) into inlined while loops with field-number dispatch. The win comes from eliminating megamorphic callback dispatch — V8 can now fully inline each generated reader.

End-to-end on the tile benchmark: write throughput improved from ~169 MB/s to ~215 MB/s (~27% faster), and read throughput from ~150 MB/s to ~187 MB/s (~25% faster).

[stepankuzmin]

Nice!

[stepankuzmin]

Impressive!

[stepankuzmin]

Great success!