fix: dispatch on type instead of length in ReferenceLoader.load_audio()

[yuvrajnode]

ReferenceLoader.load_audio() accepts bytes | str and uses len(reference_audio) > 255 as a proxy for "this is raw audio rather than a path". That heuristic is wrong in three of the four possible cases.

Current behavior

input	result
bytes shorter than 256 (truncated/empty upload)	❌ TypeError: argument should be a str object or an os.PathLike object returning str, not <class 'bytes'>
bytes longer than 255 (normal audio)	✅ works
str path that does not exist	❌ TypeError: a bytes-like object is required, not 'str'
valid str path longer than 255 chars	❌ misread as audio data → same TypeError

Only the common "long bytes" case behaves as intended. Short byte strings fall through to Path(bytes), and any str that fails the .exists() check reaches io.BytesIO(str) — neither accepts the type it is handed.

Impact

Both production callers reach this through VQManager.encode_reference() with bytes — audio_to_bytes() for id-based references, and ServeReferenceAudio.audio for API requests. So a small or malformed upload surfaces an opaque TypeError about os.PathLike rather than a meaningful decode error.

Fix

if isinstance(reference_audio, bytes):
    reference_audio = io.BytesIO(reference_audio)

Dispatch on the actual type and let torchaudio.load() report missing or undecodable files — it already does so with a clear message. This fixes all three broken cases and leaves the working one unchanged.

The repo has no test suite, so I kept this to the behavioral fix rather than introducing pytest as a new dependency.

[yuvrajnode]

Hi @tongtong-lu, @leng-yue, @Stardust-minus, @Whale-Dolphin

Could you please take a look when you have a moment? Let me know if you need any additional context or if any changes are required. Thanks!