Accelerate 1.15.0 fixes activation checkpointing that saved too much

Hugging Face shipped accelerate 1.15.0 on September 9, fixing an activation checkpointing bug specific to FSDP2. The wrapping logic was applying checkpointing to self_attn, mlp and the norm modules individually instead of the transformer layer that contains them. That meant every activation passed between those child modules stayed resident in memory for backward, which defeats a chunk of what checkpointing is supposed to save.

PR #4172, referenced in the 1.15.0 release notes, wraps the layer itself instead of its children. The fix is narrow: it only matters if you're using FSDP2 with activation checkpointing on a model where the layer boundary and the child module boundaries don't coincide, which is most transformer implementations.

The same release adds activation_checkpointing_offload, an FSDP2-only option that moves per-layer checkpoint inputs to pinned CPU memory. Hugging Face says the gradients it produces are identical to plain checkpointing, just with the inputs parked off-GPU between forward and backward.

A second fix in this release addresses FSDP2 plus PEFT plus FULL_STATE_DICT, which was dropping every rank's adapter shard except rank 0 when saving. That's a separate bug from the checkpointing one, but anyone who's been debugging missing LoRA weights after a distributed save should check #4206.

Neither fix changes the training math. They change how much memory a correct FSDP2 setup actually uses, which is the whole point of running activation checkpointing in the first place.