Add multimodal model support (STT/TTS/image) + fix UTF-8 server crash

[caseymcc]

Extend arbiterAI beyond chat/embeddings to speech-to-text, text-to-speech, and image generation, across cloud and local engines, and fix a server-wide crash on invalid UTF-8 in model output.

Data model & routing

• New request/response types: AudioTranscription, Speech, ImageGeneration.
• BaseProvider gains non-pure transcribe/synthesizeSpeech/generateImage (default NotImplemented); ArbiterAI dispatch mirrors getEmbeddings.
• modes:: constants; ModelInfo.mode is the router key.

Cloud path

• OpenAI provider implements transcription (multipart), speech (binary out), and image generation (url/b64_json).
• Server endpoints: POST /v1/audio/transcriptions, /v1/audio/speech, /v1/images/generations, with requireModelMode() enforcing mode.
• Modality-aware pricing (per-image / per-audio-second / per-character).

Local engines (default-OFF CMake flags; vcpkg overlay ports)

• whisper.cpp (STT), stable-diffusion.cpp (image, PNG+base64 via zlib), vibevoice.cpp (TTS). Each overlay port strips its bundled ggml and links the llama port's ggml so exactly one ggml exists.
• Orpheus TTS via chatllm.cpp: because chatllm bundles a forked ggml and its C binding exposes no audio, it is isolated behind a self-contained engine .so (orpheus-engine port) that exports arbiter_orpheus_tts; the Orpheus provider dlopens it (RTLD_LOCAL) so the forked ggml never clashes with the host llama ggml.
• Mock provider implements all three modalities for tests.

Feature requests (docs/feature_request.md)

• Auto-download: BaseProvider::resolveDownloadableModelFile() fetches local weights from variants[].download.url to the StorageManager models dir (multi-file bundles supported); whisper/sd/vibevoice/orpheus delegate to it.
• Modality speed telemetry: InferenceStats carries modality/images/steps/ audio-seconds/realtime-factor/steps-per-second/cost, computed at dispatch and exposed per-entry on /api/stats/history and aggregated on /api/stats.
• Schema: whisper_options / sd_options blocks (additionalProperties kept false); parsed into ModelInfo and applied as provider defaults.

Bug fix: UTF-8 crash (server-wide SIGABRT)

• safeDump() serializes every model-text response body with error_handler_t::replace + try/catch (streaming deltas, non-streaming chat, transcription/image) so no model output can abort the process.
• Root cause: Utf8StreamBuffer in the llama provider's generation loops holds an incomplete multi-byte tail until the next token completes it, so streaming deltas never emit a split character.

Tests: provider unit tests (mock/whisper/sd/vibevoice/orpheus), modality dispatch + pricing, telemetry aggregation. Full suite green (local). Design + deployment notes in docs/tasks/multimodal_model_support.md.

Claude-Session: https://claude.ai/code/session_014DA3HMkstTTLtakWda7em4