feat: Phase 2 — Text Intelligence (/read), models endpoints, Stream ergonomics

[dg-coreylweathers]

Phase 2 — Text Intelligence & stream ergonomics

Stacked on top of #164 (Phase 1). Base is feat/phase-1-quick-wins, so this diff shows only Phase 2; GitHub will retarget it to main once #164 merges.

What's included

• Text Intelligence - Analyze (Text to Text) #94 — Text Intelligence (/v1/read). Deepgram::text_intelligence() → Read with analyze_text / analyze_url, an Options builder (sentiment, summarize [boolean, not v2], topics, intents, custom topics/intents + modes, language, tag), and a response module reusing the shared sentiment/topic/intent types with a Read-specific Summary { text }. Example analyze_text.
• Models-listing endpoints. Deepgram::models() → Models with get_models, get_model, get_project_models, get_project_model; typed ModelsResponse / Model / ModelMetadata. Example models.
• [Enhancement] Add async Stream trait implementation for streaming transcription result consumption #163 — futures::Stream. TranscriptionStream already implements Stream<Item = Result<StreamResponse>>; re-exported at the crate root and documented, with a new stream_futures example showing StreamExt combinators. Verified against the current code (src/listen/websocket.rs:793).

Verification (Rust 1.97 container)

• build / clippy -D warnings / fmt / cargo test --all --all-features (139 tests, incl. new /read + models deserialization tests) ✅
• per-feature cargo check (no-default / listen / speak / manage) ✅
• cargo semver-checks: no new breaking changes — the only 2 breaks are the Phase 1 Extra metadata support #130 streaming ones inherited from the base branch.

Closes #94, #163

[dg-coreylweathers]

Live-verified against the production API (Rust 1.97 container):

• analyze_text → /v1/read: sentiment/summary/topics/intents deserialized correctly.
• models → /v1/models: full STT/TTS list parsed, including TTS metadata.accent.
• stream_futures → STT WebSocket: StreamExt::filter_map extracted final transcripts.

Also addressed the review NIT: Read's inner field is now private (matching Models/Scopes), dropping the misleading #[allow(unused)].