Switch back to cuCascade main during development cycle

[pentschev]

#1145 pinned to NVIDIA/cuCascade@d515bb0 for two reasons:

1. Pinning is always required before a RAPIDS release;
2. NVIDIA/cuCascade@8a3b42f broke RapidsMPF.

This unpins cuCascade now that the issue has been fixed.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[pentschev]

/ok to test

[pentschev]

Current cuCascade main NVIDIA/cuCascade@76a5966 still doesn't fix the original issue NVIDIA/cuCascade@8a3b42f introduced. This state was confirmed with a nightly image with rebuilt rapidsmpf+(lib)cudf_streaming+cudf_polars in a 2xNVL72 nodes:

RuntimeError: Multiple ranks share the same GPU (UUID collision detected). UUIDs: ['5961bd98-29e9-6744-49bb-e2e2b7a7b79a', '5961bd98-29e9-6744-49bb-e2e2b7a7b79a', '5961bd98-29e9-6744-49bb-e2e2b7a7b79a', '5961bd98-29e9-6744-49bb-e2e2b7a7b79a', 'f53cc9a3-3fa5-5ec2-6d0a-b9a933a737de', 'f53cc9a3-3fa5-5ec2-6d0a-b9a933a737de', 'f53cc9a3-3fa5-5ec2-6d0a-b9a933a737de', 'f53cc9a3-3fa5-5ec2-6d0a-b9a933a737de']. Set allow_gpu_sharing=True to allow this.

[pentschev]

The issue above should be fixed by NVIDIA/cuCascade#191, once that is merge we should then be able to safely do the same to this PR.