feat: SCRIBERR_ENABLED_MODELS to skip initializing unused models

[Swahjak]

Closes #473

Motivation

On startup ModelRegistry.InitializeModels calls PrepareEnvironment on every registered adapter, regardless of which model the user actually runs. For a setup that only transcribes through an OpenAI-compatible endpoint, first boot still installs the Python environments (uv sync, torch/CUDA) and downloads the weights for WhisperX, Parakeet, Canary, Voxtral, PyAnnote and Sortformer — several GB of disk and a long wait for models that will never be used.

What this adds

A new environment variable, SCRIBERR_ENABLED_MODELS, holding a comma separated list of model IDs:

# cloud transcription only, no local model environments
SCRIBERR_ENABLED_MODELS=openai_whisper

# local WhisperX with PyAnnote diarization
SCRIBERR_ENABLED_MODELS=whisperx,pyannote

Valid IDs are the registered adapter IDs: whisperx, parakeet, canary, voxtral, openai_whisper, pyannote, sortformer.

Behaviour

• Unset or empty: nothing changes. Every registered model is prepared on startup exactly as today, so existing installs are unaffected.
• When set, InitializeModels only calls PrepareEnvironment on adapters whose ID is in the list. Skipped adapters get no uv sync, no weight download, no PrepareEnvironment call at all — just a log line saying they were skipped.
• Skipped adapters stay registered. They remain resolvable through GetTranscriptionAdapter / GetDiarizationAdapter and stay listed in the UI and API, so nothing breaks if a model is looked up by capability.
• To keep a skipped model from failing when someone does select it, EnsureModelReady(ctx, modelID) prepares such a model on demand right before the job runs it (once, guarded by a mutex so concurrent jobs cannot install the same environment twice). For models that were initialized at startup, and whenever the variable is unset, this is a no-op.

Implementation

• internal/config: parses SCRIBERR_ENABLED_MODELS into Config.EnabledModels alongside the other env vars.
• cmd/server/main.go: passes it to the registry in registerAdapters.
• internal/transcription/registry: SetEnabledModels, the init-time filter, and EnsureModelReady for on-demand preparation.
• internal/transcription/unified_service.go: calls EnsureModelReady before transcription and before separate diarization.
• README: new row in the environment variable table plus a short section explaining the trade-off.

Testing

• go build ./... and go vet ./... clean.
• New internal/transcription/registry/registry_test.go covers the default (all models prepared), the filtered case (disabled model never prepared but still registered), and on-demand preparation happening exactly once.
• go test ./internal/... ./tests/... passes apart from TestListTranscriptionJobsDeltaSync, which fails on unmodified main as well.
• The Docker-based model download path was not exercised locally.