Without --convert, whisper-server fails to process any input file

Run `whisper-server -m ggml-small.en.bin --host 127.0.0.1 --port 8091` and then use any file. Transcription fails, but it looks like ffmpeg is receiving part of the input file's binary data as filename. Logs: ``` whisper server listening at http://127.0.0.1:8091 Received request: recording-20141111223317.wav ffmpeg_decode_audio: RIFF5 Couldn't open input file RIFF5 error: failed to ffmpeg decode 'RIFF5' error: failed to read audio data Received request: signal-2026-04-25-21-09-37-543.m4a Couldn't open input file error: failed to ffmpeg decode '' error: failed to read audio data ``` Call tree (on fc674574ca27cac59a15e5b22a09b9d9ad62aafe): - `examples/server/server.cpp#L859` calls read_audio_data passing contents as first parameter - `read_audio_data` expects a **filename** as first parameter - `read_audio_data` passes this to ffmpeg_decode_audio - `ffmpeg_decode_audio` uses the actual file contents as filename, always fails.