Allow large-v3-turbo as a WhisperX model

[Rediwed]

The bundled faster-whisper already resolves large-v3-turbo, but the WhisperX adapter's parameter schema hard-codes the allowed model names, so a job accepted by the API fails a second later in the worker:

single-track processing failed: transcription failed: invalid parameters:
invalid value for parameter model: value large-v3-turbo not in allowed options:
[tiny tiny.en base base.en small small.en medium medium.en large large-v1 large-v2 large-v3]

Adding the name to the two lists in whisperx_adapter.go and to WHISPER_MODELS in the config dialog is all that is needed; nothing else changes.

Why it is worth having

Measured CPU-only on a low-end 6-core x86 machine, int8, 4 threads, batch size 4, warm model cache, on an 18-minute recording. Both runs went through the same harness, so the model is the only variable:

Model	Wall clock
large-v3	1611 s
large-v3-turbo	1036 s (1.55x)

Accuracy held up in that sample. On proper nouns turbo was in fact the steadier of the two: for one place name it settled on a single spelling in 8 of 9 occurrences, where large-v3 produced five different spellings of the same word. Total word count differed by about 4%.

Related

#265 proposes the distil models. I left those out here on purpose: Distil-Whisper is English-only, while large-v3-turbo is multilingual, so it is the safer one to expose by default. The two changes are independent and could both land.