[tisonkun]

Adds ThetaUnion::serialize(ordered) and TupleUnion::serialize(ordered) so callers can serialize union results directly. Tuple serialization borrows retained summaries, avoiding the clones and associated allocations required by to_sketch(ordered).serialize().

The shared union code uses explicit owned and borrowed entry collection, with shared theta filtering, nominal-capacity trimming, and ordering. Each family's compact encoder lives in its serialization module and serves both compact sketches and unions. Theta output is byte-identical to to_sketch(ordered).serialize(); Tuple output is byte-identical when ordered and semantically equivalent when unordered, where entry order may differ.

Combines #322 and #333 by @Aschenbath, retaining contributor attribution in the commit.

Closes #320.
Closes #321.

Validation

• cargo x prepare-testdata with the pinned C++, Go, and Java fixtures.
• cargo x check across the feature matrix.
• cargo x test: 812 tests, including doctests, passed.
• cargo x lint and cargo +1.86.0 check -p datasketches --all-features passed.
• cargo x bench -- serde --test passed, including all 40 new Theta/Tuple implementation cases.
• Serialization coverage includes fresh/reset unions, exact and estimation modes, trimming, sampled inputs, non-empty results retaining no entries, custom seeds, summary clone counts, and subsequent merges of restored Tuple results.

Performance

Measured both paths at 0a287e2d2d296cdfb4a45d9e08430942452e154d on Apple M4 Max, macOS 26.6.2, Rust 1.101.0-nightly (8006091eb). Commands: cargo x bench -- theta::serde --raw and cargo x bench -- tuple::serde --raw. Each scenario reuses its union for the registered number of serializations; timing includes fixture construction and output destruction. Results below are median milliseconds and min–max over five interleaved repetitions.

Workload	Rounds	Via compact sketch	Direct
Theta, unordered, 40 inputs	1,000,000	237 (233–249)	234 (229–248)
Theta, ordered, 8,000 inputs	10,000	323 (321–323)	321 (321–324)
Tuple labels, unordered, 100	20,000	244 (243–249)	44 (43–46)
Tuple labels, ordered, 80,000	500	379 (375–385)	73 (72–73)
Tuple u64, unordered, 80,000	5,000	124 (123–129)	115 (115–117)
Tuple u64, ordered, 80,000	5,000	251 (249–251)	272 (270–277)

The labels summary holds four strings per key. Its non-empty scenarios take roughly 17–29% of the compact-sketch path's time. Scalar Tuple results are mixed, including a roughly 9% slowdown in the ordered 80,000-input case. Theta ranges overlap. These measurements support the heap-backed-summary use case, not a general serialization speedup.