perf: trust Uint8Array.fromBase64 to check non-whitespace chars

[ChALkeR]

If we are trusting the decoder at all, there is no reason to not trust it to thow on non-alphabet chars

The only place where spec behavior differs from the behavior wanted by this lib is ASCII whitespace

Before:

base64 (decode) 32 B scure x 2,881,844 ops/sec @ 347ns/op ± 7.01% (208ns..12ms)
base64 (decode) 64 B scure x 2,724,795 ops/sec @ 367ns/op ± 7.23% (208ns..11ms)
base64 (decode) 1 KB scure x 836,820 ops/sec @ 1μs/op ± 3.91% (958ns..7ms)
base64 (decode) 8 KB scure x 137,193 ops/sec @ 7μs/op
base64 (decode) 1 MB scure x 1,172 ops/sec @ 853μs/op

After:

base64 (decode) 32 B scure x 3,058,103 ops/sec @ 327ns/op ± 8.23% (166ns..17ms)
base64 (decode) 64 B scure x 3,184,713 ops/sec @ 314ns/op ± 8.32% (166ns..17ms)
base64 (decode) 1 KB scure x 1,923,076 ops/sec @ 520ns/op ± 7.34% (291ns..13ms)
base64 (decode) 8 KB scure x 535,045 ops/sec @ 1μs/op ± 2.80% (1μs..3ms)
base64 (decode) 1 MB scure x 6,576 ops/sec @ 152μs/op

Tests results are from Node.js 25 nightly, but behavior on other engines with native fromBase64 is similar

See:

• https://tc39.es/proposal-arraybuffer-base64/spec/#sec-uint8array.frombase64
• https://tc39.es/proposal-arraybuffer-base64/spec/#sec-frombase64
• https://tc39.es/proposal-arraybuffer-base64/spec/#sec-skipasciiwhitespace

---

Note

Invalid chars will now throw a SyntaxError instead of an Error
That's already the case for e.g. a==a though

If you want to normalize errors, that is doable with a try-catch