Several iOS improvements

[pcuenca]

• Dramatic performance improvements (~960ms to ~30ms); also submitted as #23.
• Load model once.
• Update output as generation takes place.
• Report ms/token when done.

[pcuenca]

@NouamaneTazi you can merge it if you want, and we can work on model download in a new PR, if necessary.