feat(diarizer): expose per-chunk embeddings on DiarizationResult

[adamsro]

Why is this change needed?

Surfaces the per-chunk speaker embeddings the offline pipeline already computes internally (currently only reachable via --export-embeddings JSON dump in the CLI), so consumers can implement chunk-granularity post-processing without re-running the embedding model.

The motivating use case is downstream cluster-purity correction: the clustering step occasionally lands a fraction of one speaker's chunks in another speaker's cluster (mega-cluster contamination on long files, bidirectional smearing between similar voices). With per-chunk embeddings plus their cluster assignments exposed on DiarizationResult, downstream code can compute per-cluster centroids and migrate chunks whose own-cluster cosine is dominated by an alternate centroid — pure NumPy/BLAS-equivalent post-processing, no extra model calls.

What changed

• New public ChunkEmbedding struct (Sendable, Codable) carrying speakerId, chunkIndex, speakerIndex, startTimeSeconds, endTimeSeconds, embedding256, and rho128. Speaker IDs follow the same "S\(cluster + 1)" convention as TimedSpeakerSegment.speakerId so chunk embeddings align to segments by string equality. rho128 is non-optional and empty when no PLDA model is loaded, matching the internal TimedEmbedding.rho128 shape.
• New optional chunkEmbeddings: [ChunkEmbedding]? field on DiarizationResult (defaults to nil, populated only when opted in).
• New OfflineDiarizerConfig.exposeChunkEmbeddings: Bool flag (defaults to false). When enabled, OfflineDiarizerManager.process(...) maps the internal [TimedEmbedding] + assignments to the public [ChunkEmbedding] array via buildPublicChunkEmbeddings(...).
• 9 new unit tests in OfflineModuleTests.swift covering default values, opt-in behavior, DiarizationResult initializer round-trip, field round-trip, the cluster-int → "S\(N)" mapping, length-mismatch handling (returns empty + warns), empty input, and Codable round-tripping.

Backwards compatibility & performance

• Fully opt-in. With `exposeChunkEmbeddings = false` (the default), the new code path is one boolean check that lands in the `nil` branch — no extra allocation, no extra compute, no memory cost.
• `DiarizationResult.init(...)` adds `chunkEmbeddings:` between `speakerDatabase:` and `timings:` with default `nil`. All existing callers in this repo use named arguments (`DiarizerManager.swift:220`, with one positional-`segments`-only call at `:223`), so no call site changes.
• `OfflineDiarizerConfig.init(...)` adds `exposeChunkEmbeddings:` with default `false`. The community/community-1 presets and existing init overloads keep their behavior.
• When opted in: ~1–2 MB per hour of audio for the embedding + PLDA payload. Helper is O(n) over already-computed data — no model calls, no audio access, ~1 ms even on 90-min files.

Tests & lint

```
$ swift test
Executed 1352 tests, with 24 tests skipped and 0 failures (0 unexpected)

$ swift format lint --recursive --configuration .swift-format Sources/ Tests/

(only pre-existing warnings on Fa/Fb in OfflineDiarizerConfig and

Sortformer/CLI files; none introduced by this change)

```

Out of scope

The downstream chunk-cluster migration consumer that motivated this API lives outside this repo and is not part of this PR.