simplify meta_init (rope embeddings)

[lessw2020]

Currently we cannot use the simpler with torch.device('meta') and a lambda for FSDP param_init_fn. Rather we have a pre and post handler to move the buffers and modules
Depending on the variation, you will hit a couple different errors but for the error currently blocking the simple meta init, the error is:

NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

This looks like we need to implement the equivalent of reset_params() for the rope embeddings class.
Making an issue to resolve this so we can get the working meta init landed and then come back to improve it by resolving this issue.

[lessw2020]

it seems others have solved this by simply keeping rope embeddings completely out of the whole meta init process:
https://github.com/allenai/OLMo/blob/cb711e24538983c9a1bbdc12f80faccb8f67b758/olmo/model.py#L283

[awgu]

I think we can close this since FSDP2 addresses this problem differently.