Fix Base36 round-trip for empty, all-zero, and high-bit-set inputs (fixes #67)

[youdie006]

What

Base36.encode/decode route bytes through BigInteger, which loses fidelity two ways, so decode(encode(x)) does not return x:

1. Empty / all-zero (reported in Base36 fails round trip encoding of empty byte arrays #67). new BigInteger(1, allZeros).toString(36) yields "0" not "", and new BigInteger("0", 36).toByteArray() yields [0] not an empty array. Each side adds a spurious byte/char and they compound over a round trip: [] -> "0" -> [0, 0].
2. Sign byte (latent, not in the report). For a value whose top byte has the high bit set, BigInteger.toByteArray() prepends a 0x00 two's-complement sign byte that decode never strips: [0x80] -> "3k" -> [0x00, 0x80]. This affects roughly half of all inputs.

Both are reachable through the public Multibase API, which routes the Base36/Base36Upper codes (k/K) straight to these methods.

Fix

Mirror the contract of this repo's own Base58 (empty -> empty, leading zeros preserved, no sign byte), keeping the existing leading-zero-prefix logic:

• encode: emit "" for a zero value instead of "0".
• decode: return new byte[0] for empty input, and strip a single leading 0x00 sign byte from toByteArray() for a non-zero value.

For a positive BigInteger, toByteArray() has at most one leading 0x00 (the sign byte), and only when the top magnitude byte's high bit is set, so stripping one is always safe; genuine leading-zero bytes of the input are still carried by the separate zeroPrefixLength path.

Tests

New Base36Test (JUnit 5, matching the existing suite):

• size-preserving round trip for {}, {0}, {0,0}, {0x80}, {0x01};
• empty <-> "";
• the pre-existing multibase base36 vectors stay byte-identical, and the leading-zero vector still decodes to 0x00 ++ decode(body);
• an exhaustive round trip over every array of length 0..2.

The exhaustive check goes from 33027 of 65793 inputs failing (before) to 0 (after). Pre-existing MultibaseTest (108) and MultibaseBadInputsTest (9) stay green; 121 tests pass total.

Thanks to @hossman for the clear report and repro.

---

This change was prepared with AI assistance and reviewed by me before submission.