Avoid default MP3 transcoding for transcription benchmarks

[ushaket]

Problem Statement

When benchmarking transcription/translation workloads, GuideLLM currently transcodes input audio to MP3 by default through the encode_media preprocessing path (encode_audio(audio_format="mp3") default). This changes the original dataset audio format before request submission.

Proposed Solution

Change default audio encoding behavior to avoid forced MP3 transcoding.

Suggested behavior:

1. If audio_format is explicitly provided by the user, use it.
2. Otherwise, infer format from source metadata/path (file suffix, URL suffix, or dataset-provided format).
3. If format cannot be inferred, default to wav (safe/common fallback), not mp3.

Optional:

• add a warning/log when fallback-to-WAV is used due to missing format metadata,
• document this behavior in benchmarking docs and CLI examples.

Alternatives Considered

Always force WAV instead of MP3: better than MP3 for fidelity, but still ignores source format unnecessarily.

Usage Examples

Additional Context

No response