Avoid keeping redundant copies of model weights in memory during load

[drdaxxy]

Before this change, whisper.load_model keeps the whole weights file in scope until after tensors are moved to the device.

This might not seem important for sample code, but this change makes it possible to load the large model on a free Colab instance, which otherwise runs out of memory (inference afterwards is no problem)!

[nlgtuankiet]

I can confirm that it works!

Before: using large model on a free Colab instance -> Out of memory error
After: using large model on a free Colab instance -> Run successfully!

I'm not sure if these changes affect the output quality, can someone make a comparison to make sure?

[jongwook]

This is great! I wasn't thinking about Colab memory limits.

Let me make some cosmetic changes, test on my side, and then merge. Thanks!