Init weights only if not loading a checkpoint

[carmocca]

Initializing the weights is wasteful if a checkpoint will be loaded.

This assumes that checkpoints are complete (loadable in strict=True mode). Let me know if that's too strong of an assumption generally.

It might also be problematic with non_persisent=True buffers, if they were supported (#316). That is if you expect to initialize them in init_weights. Although one could argue for a separate init_buffers() method

[awgu]

I agree with the spirit of this change :)

init_buffers for non-persistent buffers makes sense to me too.

[tianyu-l]

cc: @fegin