feat(speech): add OmniRoute text-to-speech provider

[ekinnee]

Summary

• add the OmniRoute registerSpeechProvider adapter for POST /v1/audio/speech
• support explicit model selection, OpenAI-compatible voices, mp3/opus/wav, provider-specific auth/base URL overrides, and guarded bounded binary responses
• document speech ownership/configuration and extend manifest/compatibility probes/tests/generated dist/

Related issue

Closes #15

Validation

• pnpm test (143 tests)
• pnpm build
• git diff --check
• node .github/scripts/upstream-compatibility-probe.mjs --source-contract
• packed dist/index.js registration probe with openclaw@2026.7.1 (provider/modelCatalog/embedding/image/speech/video/music/webSearch all registered once)
• direct OmniRoute catalog and speech-route probe performed without exposing credentials or persisting audio

Notes for reviewers

• No credentials, private endpoints, or production-only data included
• Capability and manifest ownership changes are documented
• Explicit speech models are required; the plugin does not synthesize audio catalog defaults.
• Response validation requires matching MIME/signature, including an Ogg OpusHead for Opus output.
• The live catalog exposed Ollama Cloud chat/reasoning/vision models but no speech-capable model. The advertised Cloudflare MeloTTS entry was rejected by OmniRoute's speech router as an invalid speech model, so no live audio result is claimed.