Allow LPLayerNorm and LPGroupNorm to support self.bias or self.weight = None

[abhi-mosaic]

We are experimenting with removing all biases from our MosaicGPT models. When we do so with torch.nn.LayerNorm it works, but with LPLayerNorm it fails. This PR:
(1) Modifies the weight copying from torch.nn.LayerNorm to LPLayerNorm that occurs during module surgery to check for None types.
(2) Adds a check in the forward() method if self.bias is None or self.weight is None, and doesn't downcast the None parameters.
(3) Updates a test to run on a model where some LayerNorms have both weights and biases, some have self.bias = None, and some have self.bias = None and self.weight = None.

This PR also does (1), (2), and (3) for LPGroupNorm.

[abhi-mosaic]

@bandish-shah there is a moderate likelihood that our MosaicGPT in examples will need this feature, so until this patch is released in Composer, we likely cannot tag a stable release of Examples against a stable release of Composer (or we will need to maintain a second branch or something like that)

Basically this might be worth an 0.13.2, if you have other stuff you want to get in there.

[mvpatel2000]

@abhi-mosaic Can you please add a unit test ensuring this works? Similarly, can you please add the same fix and unit test to LowPrecisionGroupNorm? it's probably something simple like extending existing models they use in unit tests to have a flag for bias or not (with default on).

Basically this might be worth an 0.13.2, if you have other stuff you want to get in there.

We do have a lot of checkpointing fixes coming in this week....

[bandish-shah]

Basically this might be worth an 0.13.2, if you have other stuff you want to get in there.

@abhi-mosaic let's hold off, there are other fixes in progress. We can target the 0.13.2 hot patch for late next week.

[mvpatel2000]

LGTM! Thanks for finishing this

[abhi-mosaic]

Thanks @nik-mosaic !!!