perf: rewrite src/uft8.ts using TextDecoder

[rwv]

You can rewrite src/utf8.ts using TextDecoder for better performance and robustness.

/** Converts a string to an array of UTF8 bytes. */
export function utf8Array(value: string) {
  const encoder = new TextEncoder();
  const uint8Array = encoder.encode(value);
  const buffer = new Int32Array(uint8Array.buffer);
  return buffer;
}

[Mister-Hope]

I am not pretty sure about the "better performance" part.

See https://github.com/Mister-Hope/bcrypt-ts/tree/text-encoder for code details.

I agree that TextEncoder can be performant, but it's init process of .encode() is a bit huge.

✓ __tests__/utf8.bench.ts > convertToUFT8Bytes > hello 4608ms
     name                    hz     min     max    mean     p75     p99    p995    p999     rme  samples
   · TextEncoder   2,558,457.74  0.0002  0.4626  0.0004  0.0004  0.0035  0.0043  0.0060  ±0.54%  1279229
   · original     19,811,685.92  0.0000  2.2516  0.0001  0.0000  0.0001  0.0001  0.0002  ±0.92%  9905843   fastest

 ✓ __tests__/utf8.bench.ts > convertToUFT8Bytes > samplePassword 3436ms
     name                    hz     min     max    mean     p75     p99    p995    p999     rme  samples
   · TextEncoder   2,461,509.85  0.0002  0.7057  0.0004  0.0004  0.0036  0.0043  0.0057  ±1.19%  1230755
   · original     12,154,916.98  0.0001  0.3119  0.0001  0.0001  0.0001  0.0002  0.0003  ±0.53%  6077459   fastest

 ✓ __tests__/utf8.bench.ts > convertToUFT8Bytes > lorem 1803ms
     name                   hz     min     max    mean     p75     p99    p995    p999     rme  samples
   · TextEncoder  1,898,518.55  0.0002  6.3016  0.0005  0.0004  0.0028  0.0040  0.0074  ±3.23%   949260   fastest
   · original     1,324,877.16  0.0006  0.5155  0.0008  0.0007  0.0011  0.0013  0.0078  ±0.72%   662439

 ✓ __tests__/utf8.bench.ts > convertToUFT8Bytes > lorem long 1463ms
     name                   hz     min     max    mean     p75     p99    p995    p999     rme  samples
   · TextEncoder  1,068,079.24  0.0003  1.2460  0.0009  0.0006  0.0072  0.0097  0.0311  ±1.29%   534040   fastest
   · original       315,078.54  0.0024  0.5745  0.0032  0.0030  0.0042  0.0077  0.0921  ±0.63%   157540

 BENCH  Summary

  original - __tests__/utf8.bench.ts > convertToUFT8Bytes > hello
    7.74x faster than TextEncoder

  original - __tests__/utf8.bench.ts > convertToUFT8Bytes > samplePassword
    4.94x faster than TextEncoder

  TextEncoder - __tests__/utf8.bench.ts > convertToUFT8Bytes > lorem
    1.43x faster than original

  TextEncoder - __tests__/utf8.bench.ts > convertToUFT8Bytes > lorem long
    3.39x faster than original

In most use cases for bcrypt-ts for me, is to deal with passwords (tokens), which I am not expecting these things to have length greater than 32.

The original approach is still much faster than TextEncoder, even I already reuse the TextEncoder instance and avoid the converting from Uint8 to Int32. So at least I think the current codes are fine in normal usage sences.

Discussion about this is still welcomed, but I will not do any change until we align our opinions.

Edited:

I believe it's possible to change to TextEncoder if the length is greater by a threshold, but I have no ideas how this should be with different devices (performance/numbers with CPU core) and different runtimes (Node.js and browsers's different V8 and other runtime with different implantations for TextEncoder)

To deal with long texts I believe you shall use C++ bcrypt