feat: improve language detection with multi-sampling

[Zonrotan]

Description

The new auto-detect works great, but has a hidden flaw: when language=None, Whisper natively limits its language detection to the first 30 seconds of the audio.

If a video starts with silent logos, long musical intros, or background noise, Whisper often hallucinates a random language (e.g., guessing Thai for a Swedish video). Because Whisper locks in its first guess for the entire file, this ruins the transcription and triggers unnecessary translation logic.

Solution

This PR bypasses Whisper's 30-second limitation with a lightweight detect_robust_language helper function that samples three snippets from the file. Instead of trusting just the first 30 seconds, it:

1. Samples three 30-second snippets (at 15%, 50%, and 85%).
2. Silently runs model.transcribe() on them.
3. Uses majority voting to determine the true language.
4. Falls back to native detection for videos under 2 minutes.

Why this approach?
Since the audio is already loaded as a numpy array, slicing it and running inference on these snippets takes only a fraction of a second and requires no external dependencies (like ffmpeg).

Changes Made

• Added detect_robust_language in whisper.py
• Injected the check before the main model.transcribe call (when no metadata is present).
• Muted the progress bar for the detection samples.

Sounds good, no? 😃