Added opportunity to transcribe audio directly from bytes

[KirillKukharev]

The transcribe_bytes() method has been added to the GigaAMASR class, which performs audio transcription directly from bytes in memory without file I/O.

Why is this important?

Performance and latency
Fixed file I/O: temporary files are not created
Eliminated overhead from ffmpeg subprocess call
Direct conversion of PCM16 bytes to tensors via torch.frombuffer()

Technical details
Uses load_audio_from_bytes() instead of load_audio() for direct byte conversion
Usage example:
Version with temp file (before pull request was created):

result = model.transcribe("audio.wav")

Version without file:

audio_bytes = receive_audio_from_network()  # abstract method; receive PCM16 16kHz
result = model.transcribe_bytes(audio_bytes)