Add OpenAI-compatible speech transcription endpoint

[ivan-digital]

Summary

• add POST /v1/audio/transcriptions to speech-server with the standard multipart request shape
• decode PCM16/24/32 and Float32 WAV, downmix and resample to 16 kHz, then lazily load the bundled Parakeet ONNX backend
• support json, text, verbose_json, srt, and vtt responses plus the existing bearer authentication envelope
• preserve the 1 MiB TTS limit while bounding STT uploads to 25 MiB and 10 minutes and serializing decode/inference memory
• document both HTTP audio endpoints, update package smoke checks, and keep all README language links synchronized
• repair the dependency-free Parakeet mel-contract test registration/linkage exposed by a clean configure

Architecture and risk

The HTTP layer remains model-independent through a transcription callback. The production sidecar owns model selection and lazy Parakeet initialization, so existing TTS startup time and resident memory do not change. TTS and STT share the existing inference lock. The main compatibility limit is deliberate: uploads must be uncompressed RIFF/WAVE because release packages do not ship a codec runtime.

Validation

• dependency-free CTest suite: 25/25 passed
• HTTP contract regression suite: passed
• AddressSanitizer and UndefinedBehaviorSanitizer HTTP suite: passed
• ONNX-enabled speech-server and endpoint E2E target: compiled and linked with ONNX Runtime 1.19.0
• model-backed E2E target: verified to skip cleanly when SPEECH_MODEL_DIR is unavailable
• actionlint on CI and release workflows: passed
• shellcheck scripts/speech-dispatch.sh: passed