Release the model from VRAM

[yumianhuli1]

I tried to clean the faster whisper model from the VRAM for use by other module. I del the fasterWhisper_Transcriber object and torch.cuda.empty_cache(), but there was always 200-300M of memory left uncleaned