[Aschenbath]

Closes #320. Related to the Theta comparison suggested in #322, with a separate branch based on main.

Adds ThetaUnion::serialize(ordered) for uncompressed compact output. It reuses the existing union result preparation and shares the compact encoder. The encoder can also serve a future intersection or mutable-compaction path once that operation produces canonical state. An uninitialized intersection would still need to be handled before encoding.

• cargo x test: 815 tests passed with the pinned TCK fixtures.
• cargo x check and all eight benchmark smoke cases passed.
• Compared against to_sketch(ordered).serialize() on Linux x86_64, Rust 1.99.1, via cargo x bench -- theta::serde --raw. Five repetitions, medians in milliseconds (via compact → direct): 40 inputs unordered 57 → 46, ordered 83 → 70, 8,000 inputs unordered 86 → 87, ordered 225 → 223. The measured ranges overlap.
• An isolated counting allocator measured the same allocation/reallocation counts and requested bytes for both paths: 6 calls/1,328 bytes at 40 inputs and 13 calls/163,832 bytes at 8,000 inputs for either ordering. Result preparation still uses the existing scratch vector. These measurements do not establish a general speed or allocation improvement.

AI-assisted with Codex.