By default let HF download handle cache path for WhisperKitPro Engine

[EduardoPach]

What does this PR do?

Problem

Pipeline aliases generate a new download path on each CLI run, bypassing the cache and causing unnecessary model re-downloads.

Solution

Delegate model caching to HuggingFace's built-in cache mechanism:

• Downloading: When using repo_id + model_variant, models are stored in HF's default cache path, ensuring cache hits across runs
• Local models: When model_dir is provided, models are loaded directly from that path (no download)

Changes

• Extended model downloads to WhisperKitProOrchestrationPipeline
• Added repo_id and model_variant for HF-managed downloads
• Switched models_cache_dir logic to a simple model_dir to load models from a local path directly
• Updated build_pipeline in transcription and orchestration to pass new fields to WhisperKitProConfig

[dbrkn]

lgtm