add faster-whisper as an alternative transcription backend

[ucscode]

Currently, WhisperAudioTranscriber relies exclusively on the vanilla openai-whisper package. While accurate, it runs relatively slow on standard CPU environments and uses a larger VRAM/RAM footprint compared to optimized execution backends.

I would like to propose adding faster-whisper (built on CTranslate2) as an optional or alternative backend implementation.

Benefits

• Speed: Up to 4x faster execution compared to vanilla OpenAI Whisper at the same precision level.
• Efficiency: Uses integer quantization (int8, float16), allowing models like base and small to run exceptionally fast and light on host CPUs without crashing or swapping memory.

Proposed Implementation Details

Because faster-whisper returns an iterable generator of segment objects rather than a static dictionary payload, the transcribe() loop needs to unpack the generator tuple and read attributes natively rather than accessing keys:

segments_generator, info = self._get_model().transcribe(
    audio_path,
    word_timestamps=True,
    language=self._language,
    initial_prompt=self._initial_prompt,
)

for segment_info in segments_generator:
    segment_start = float(segment_info.start)
    # Map attributes down into Document -> Segment -> Line -> Word structures...

We could either introduce a backend="vanilla" | "faster" flag in the transcriber configuration or split them into explicit WhisperAudioTranscriber and FasterWhisperAudioTranscriber classes.