perf: use String.fromCharCode.apply in floats32ToBase64 example

[rdgordon-index]

Description:
The current encode example accumulates a binary string by appending one character per byte in a for...of loop. Each iteration allocates a new string, making the total cost O(n) allocations for an n-byte vector.

String.fromCharCode.apply(null, bytes) processes all bytes in a single native call, avoiding the per-byte re-allocations. Benchmarked with Benchmark.js against normalized unit-sphere float32 vectors across realistic embedding dimensions:

Dimension	for…of (ops/sec)	apply (ops/sec)	speedup
spec example (10 floats, 40 B)	887,207	1,473,359	×1.7
128-dim (512 B)	124,262	255,052	×2.1
256-dim (1 KB)	48,772	132,912	×2.7
768-dim / BERT-base (3 KB)	22,171	45,845	×2.1
1536-dim / OpenAI ada (6 KB)	10,725	24,995	×2.3
3072-dim / OpenAI large (12 KB)	5,167	10,920	×2.1

The spec example row was reproduced with:

const Benchmark = require('benchmark');
const floats = [1.2345678, -2.5, 0.0, 3.1415927, 12345.678, -0.00012345, 42.42, -999.999, 0.000001, 987654.25];
const buf = new ArrayBuffer(floats.length * 4);
const view = new DataView(buf);
floats.forEach((x, i) => view.setFloat32(i * 4, x, true));
const bytes = new Uint8Array(buf);
new Benchmark.Suite()
  .add('for…of', () => { let s=''; for (const b of bytes) s += String.fromCharCode(b); btoa(s); })
  .add('apply',  () => btoa(String.fromCharCode.apply(null, bytes)))
  .on('cycle', e => console.log(String(e.target)))
  .run({ async: false });

Works in all modern browsers with no caveats. A comment in the code notes that very large inputs should be chunked into ~8192-byte blocks to stay within engine argument count limits, though this is not a concern for any embedding dimension currently in use.

The decode example is unchanged — there is no browser-compatible equivalent speedup for that path.