[Umbrella] Long-context SFT fixes (do not merge)

[qgallouedec]

Install-only umbrella for training on 1M+ token sequences: merges

• Add FSDP2 activation_checkpointing_offload: offload checkpointed layer inputs to pinned CPU memory #4175 (FSDP2 activation_checkpointing_offload)
• Refuse context parallelism for models with sliding-window or chunked attention layers #4177 (refuse context parallelism for sliding-window or chunked attention)

so colleagues can install one branch. Not meant to merge.

#4175 is what makes Qwen3-8B fit at 1,048,576 tokens on one 8xH100 node: without it the run OOMs. Paired with huggingface/transformers#48163 the step time is 373 s at 56.2 GB per GPU, against 582 s for Qwen3-4B (the largest that fits) on unmodified main.

Used by the 1M-token context-parallelism example in huggingface/trl#6846.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.