Faster-whisper issue with the latest NVIDIA 55x series drivers

[pasinduranjula]

Faster-whisper on CUDA doesn't properly terminate with the latest NVIDIA 551 and 552 drivers on Windows. It hangs for a while and triggers an automatic Microsoft crash submission before terminating. The issues occurs with both CUDA 11 and 12, but works without issues on all previous NVIDIA drivers (e.g. 546).

A simple code like the following can trigger the issue:

`from faster_whisper import WhisperModel

model = WhisperModel("modelpath", device="cuda", compute_type="int8_float16")`

The issue seems to trigger when releasing the CUDA memory allocated for the model.

Tested with multiple faster-whisper versions (both CUDA 11 and 12).

[pasinduranjula]

The issue seems to be in the ctranslate2 library. There are some CUDA API calls in the destructors that cause the issue. https://github.com/OpenNMT/CTranslate2/blob/v4.2.1/src/cuda/utils.cc