perf: fast native UTF-16 source copy into Wasm memory on Node

[guybedford]

Copying the JS source into Wasm linear memory is ~25% of total parse time, and it was done with a per-character charCodeAt loop (~1.2 GB/s). On Node we can instead encode the string as UTF-16 straight into memory natively via Buffer (~15 GB/s, ~13x), with no other lexer changes since the lexer already indexes UTF-16 code units.

Warm parse of the full sample corpus (3057 KiB, avg of 25 runs):

• before: ~11ms
• after: ~8ms (~1.35x, ~27% faster end-to-end)

// before
copyLE(source, new Uint16Array(wasm.memory.buffer, addr, len));
// after
Buffer.from(wasm.memory.buffer, addr, len * 2).write(source, 'utf16le');

The whole gain comes from a single capability: encoding a JS string as UTF-16 directly into linear memory. Node exposes it natively; the browser has no equivalent (TextEncoder is UTF-8 only), so the portable charCodeAt path is retained as a fallback and the same speedup is left on the table anywhere a native UTF-16 encode-into-memory becomes available.

The fast path is gated strictly to the real Node Buffer on little-endian hosts (every platform Node runs on); browser bundles and big-endian hosts keep the existing code-unit copy. Existing tests pass unchanged, including the non-ASCII sample files.

[BridgeAR]

Nice!