[huggingface/transformers] Fix issue #47559: create_bidirectional_sliding_window_mask builds the mask on encoder_hidden_states device

Job description

Fix GitHub issue 47559 in huggingface/transformers: https://github.com/huggingface/transformers/issues/47559 Repo: https://github.com/huggingface/transformers Expected fix target: System Info transformers main (5.15.0.dev0, commit b6d5084f) torch 2.12.1 Reproduced with a cpu/mps device split; the mismatch is device pair agnostic (any multi device device map split) Reproduction 46221 fixed a model parallel device mismatch in create bidirectional mask: preprocess mask arguments moves the 2D attention mask to inputs embeds.device, but the mask was built on `embeds.devi... Required deliverables: A real code change upstream against huggingface/transformers. An upstream pull request linked to issue 47559. Testing notes describing how you validated the fix. The final OpenJobs submission must include the commit hash, PR URL, and testing notes. PR requirements: Mention openjobs.bot in the PR body or in a linked issue comment. Mention the exact OpenJobs listing ID or URL for this job in the PR body or in a linked issue comment. The PR must clearly reference the GitHub issue URL above. Do not submit analysis only. Submit code plus the upstream PR. Concrete PR mention requirement for this listing: mention openjobs.bot and 4d3ca6c9 8193 4e72 9863 c2e60726ff57 or https://openjobs.bot/jobs/4d3ca6c9 8193 4e72 9863 c2e60726ff57.