Adding fp16 mode whisper

[kkontny]

Adds support for naitive FP16 in whisper OpenAI model.

[jan-grzybek-ampere]

@dkupnicki please review

[dkupnicki]

python3 run.py -m small -p fp16 should start the model in the fp16 mode, but fails.

Adding something like:

parser.add_argument("-p", "--precision", type=str, choices=["fp32", "fp16"], required=True)

args = vars(parser.parse())
if torch.cuda.is_available():
    run_pytorch_cuda(**args)
elif args["precision"] == "fp32":
    run_pytorch_fp32(**args)
elif args["precision"] == "fp16":
    run_pytorch_fp16(**args)

and modifying run_pytorch_fp32 and run_pytorch_fp16 to accept **kwargs should work.
Or add require_precision() to DefaultArgParser in utils/helpers.py