[Feature] Enable any frame length for SP parallel

[MartinPernus]

Motivation

The current code forces users to shorten video length for sharding compatibility:

def shard_latents_across_sp(latents: torch.Tensor,
                            num_latent_t: int) -> torch.Tensor:
    sp_world_size = get_sp_world_size()
    rank_in_sp_group = get_sp_parallel_rank()
    latents = latents[:, :, :num_latent_t]
    if sp_world_size > 1:
        latents = rearrange(latents,
                            "b c (n s) h w -> b c n s h w",
                            n=sp_world_size).contiguous()
        latents = latents[:, :, rank_in_sp_group, :, :, :]
    return latents

Problem: When num_latent_t is not divisible by sp_world_size (which happens often in WAN with 21 temporal latent dim), the reshaping operation fails.

Proposed Solution:

• Preserve the entire video segment by padding the temporal dimension to make it divisible by sp_world_size
• Return a validity mask alongside the sharded latents
• Account for the mask in both training loss computation and inference to ignore padded regions

[MartinPernus]

Ah I see there is already work on this in #664