[Deepspeed zero3] lazy weights init

I'm pretty sure we need to follow up to the lazy weights init feature https://github.com/huggingface/transformers/pull/11471 and add under zero3 `deepspeed.zero.GatheredParameters` here (or inside `_init_weights`): https://github.com/huggingface/transformers/pull/11471/files#diff-6b72b98c4c2dcfc6cc606843917733f5d858374fbc22a735ff483bbc0c1e63eaR1275-R1276 plus need a test.