Run init_weights under no_grad

[carmocca]

The initializations in init_weights can create gradients. This is almost always not intended

The alternative would be to add the decorator to model.init_weights directly. Then the responsibility is moved to the model writer. I can change that if that's preferred.

[tianyu-l]

Thanks for the fix. I think it's fine, compared to having a decorator on every single init_weights() definition (reasoning: technically we probably only need it on the outermost one, but what if users would like to call it explicitly for a submodule -- they still have to use the context manager).