Fixes a regression where Buffer.prototype.toString("utf8") — and therefore node:string_decoder — produced different output than Node.js for buffers containing invalid UTF-8 bytes.

Repro (from #34930)

import { createHash } from "node:crypto";
import { StringDecoder } from "node:string_decoder";

const buffer = Buffer.from("R0ABkgEBAQE...AQE=", "base64"); // bytes: 47 40 01 92 01 01 ...
const text = new StringDecoder("utf-8").write(new Uint8Array(buffer));
console.log(createHash("sha256").update(text).digest("hex"));
// Node:   2637e4a587f78c8be9f18fec1ad0fcbe5ee9ca90e01c7c9ce61a7fe6ab7d6610
// Deno:   633b6f1cc3191287ce213370e649ed09700695a7485081914111f4bd553c8101 (before)

The lone continuation byte 0x92 was not replaced with a single U+FFFD the
way Node does, leading to a different decoded string and silent data corruption.

Root cause

This regressed in #30213, which optimized utf8Slice to go through
op_node_decode → v8::String::new_from_utf8. V8's NewFromUtf8 does not
follow the WHATWG "maximal subpart" replacement that Node.js (and the previous
decoder) used for ill-formed UTF-8.

Fix

Decode invalid input with Rust's String::from_utf8_lossy, which implements
the WHATWG maximal-subpart replacement (one U+FFFD per ill-formed
subsequence) and matches Node. Valid UTF-8 is borrowed without allocating, so
the optimized fast path is unaffected.

Added a regression test in tests/unit_node/string_decoder_test.ts covering
both StringDecoder.write and Buffer.prototype.toString("utf8").

Closes #34930

Closes denoland/divybot#505