calldata.encode silently accepts lossy JavaScript values #229

Summary

calldata.encode currently accepts two JavaScript value classes that cannot be represented losslessly:

1. Unsafe integer numbers. Number.isInteger() accepts values outside JavaScript's safe-integer range, then BigInt(data) encodes the already-rounded Number. The API also accepts bigint, so unsafe Numbers should be rejected and callers should use bigint for exact integers.
2. Strings with unpaired UTF-16 surrogates. TextEncoder replaces an unpaired surrogate with U+FFFD instead of failing. That means decode(encode(value)) !== value, and distinct map keys can collapse to the same UTF-8 bytes.

Example key collision:

new Map([
  ["\\uD800", null],
  ["\\uFFFD", true],
]);

Those are distinct JavaScript strings, but TextEncoder emits the same EF BF BD bytes for both keys. The GenVM Python reference encoder raises on an unpaired surrogate instead of replacing it.

Expected

• reject number values for which Number.isSafeInteger(value) is false and direct callers to bigint;
• reject unpaired UTF-16 surrogates in calldata string values and map keys;
• preserve normal Unicode, including valid surrogate pairs such as emoji.

This prevents silent wire-value changes before calldata reaches GenVM.