Fix parallel vae

[gty111]

related to #271

[gty111]

Related PR xdit-project/DistVAE#3

To consider DP, we add gather_broadcast_latents and is_dp_last_group in xFuserPipelineBaseWrapper.

gather_broadcast_latents:

1. create dp last group
2. gather latents from dp last group and concatenate the latents in the batch dim
3. broadcast latents

Then use default group to complete parallel vae.

When using parallel vae, only one process need to process img after vae. So is_dp_last_group is to correctly identify the process which needs to do sth after vae.

@Eigensystem

[Eigensystem]

I think it is not a elegant implementation because you combined all the dp latents into one tensor instead of using different dp groups to process different latent. It will increase the communication load.

[gty111]

I think it is not a elegant implementation because you combined all the dp latents into one tensor instead of using different dp groups to process different latent. It will increase the communication load.

It is indeed not the best implementation. Using different dp groups need to refactor DistVAE. Since parallel vae is an option, maybe we can first fix parallel vae in this PR, and further optimize it in the future.

[Eigensystem]

LGTM.