Re-enable short Whisper filler prompts for um/uh retention

[wassgha]

Summary

Stock Whisper cleans disfluencies, so Remove fillers often has nothing to cut. This restores <|startofprev|> filler prompting for Whisper Base/Small, but with the safer short prompts we learned we need:

• Short language-specific prompts (Um, uh, hmm, er, ah. and DE/FR/ES/ZH equivalents)
• Hard length cap (≤32 chars) — the long OpenAI-style prompt previously truncated multi-speaker / long-form ASR
• No I'm … openers (those seeded "I'm sorry" hallucination loops)

Parakeet is unchanged (no Whisper prompt path).

Changes

• keepFillers flag on Whisper models + whisperFillerPrompt(model, language)
• Worker uses the resolved short prompt via existing buildPromptedDecoderIds
• Regression test now requires short prompts (instead of forbidding all prompts)

Test plan

• npx tsx tests/models-test.ts
• npx tsx tests/vad-regression-test.ts
• npx tsx tests/fillers-test.ts
• Manual: transcribe a clip with audible um/uh on Whisper Base and confirm fillers appear so Remove fillers is enabled
• Manual: multi-speaker clip still keeps both speakers (no second-speaker drop)