Add the LZHUF compressor

[colemancda]

Summary

Adds LZHUF.compress — the encode side of the classic public-domain LZHUF.C the client embeds. Static analysis of the original binary recently named its copy of the encoder (EncodeLZHUFBlock 0x4ea760, with LZHUFPutCode/LZHUFDeleteNode/LZHUFUpdate matching the reference's Putcode/DeleteNode/update, used by its XFS write-back path), confirming the client ships the reference compressor, not just the decoder.

• Greedy LZSS matching over the 4096-byte ring via the reference's per-first-byte binary search tree (InsertNode/DeleteNode ported exactly, including full-match node replacement and lowest-position tie-breaking), each literal or (length, position) pair adaptive-Huffman coded.
• encodeChar lives on the same HuffmanTree the decoder evolves, and the encoder position tables are derived from the decoder's dCode/dLen rather than transcribed — the two sides are inverse by construction and cannot drift apart.
• BitWriter mirrors BitReader (Putcode semantics, including the two-byte spill path).
• Output carries no size header, matching every GunBound container (decoded size is stored out of band).

Tests

Six new tests in LZHUFCompressTests:

• Byte-for-byte identity with the reference encoder on all three .lzh fixtures (they were produced by the original LZHUF.C; a faithful port must reproduce them exactly, minus the CLI's 4-byte size header) — the strongest validation possible.
• Round trips of the sample fixtures, the real characterdata.dat payload, degenerate inputs (empty / single byte / all-identical), >32K-symbol data forcing the adaptive-Huffman reconst rebuild on both sides, and multi-window repetitive data exercising both position-code bands.

Test plan

• swift test — 566 tests passing
• CI green on this PR