perf: reduce allocation churn in operator serialization

[ntedvs]

Summary

Each Operator.toBytes() call created ~10+ intermediate Uint8Array allocations (one per operand, one per space separator, one concat result). Then serializeOperators() did the same pattern again at a higher level with newlines. For 100 rectangles, that's ~7,600 throwaway allocations per iteration.

Added Operator.writeTo(writer: ByteWriter) so operators write directly into a shared buffer instead of building intermediate arrays. Rewrote serializeOperators() to use a single ByteWriter instead of collecting per-operator byte arrays and concatenating.

Also removed byteLength() (dead code, zero production callers).

Numbers

Drawing benchmarks (absolute throughput):

Benchmark	Before	After	Speedup
draw 100 rectangles	1,528 ops/s	3,635 ops/s	2.38x
draw 100 circles	756 ops/s	2,161 ops/s	2.86x
draw 100 lines	1,787 ops/s	3,921 ops/s	2.19x
draw 100 text lines	507 ops/s	1,026 ops/s	2.02x
10 pages mixed content	532 ops/s	1,396 ops/s	2.63x

Comparison against pdf-lib (draw 50 rectangles):

	Before	After
libpdf vs pdf-lib	1.75x faster	3.56x faster

Test plan

• All existing tests pass (operators, drawing, integration)
• Typecheck clean

[vercel]

@ntedvs is attempting to deploy a commit to the mythie's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Mythie]

Great work! I had similar changes on a branch but never got back to it 🙌🏻

[ntedvs]

Thank you! Any tips for contributing going forward?

[ntedvs]

@Mythie ?