Add weights_only=True to torch.load() calls

[michaelv2]

Summary

• Adds weights_only=True to all 8 torch.load() call sites across the codebase
• Prevents arbitrary code execution via crafted .pt files (pickle deserialization attack)
• Aligns with PyTorch's recommended practice since 2.0 and addresses CVE-2025-32434 for users on PyTorch < 2.6.0 (weights_only is available since 2.0 but defaults to False, so this is no-op on 2.6+ and avoids breaking changes for 2.0-2.5)

Context

torch.load() without weights_only=True uses Python's pickle module, which can execute arbitrary code during deserialization. Since all from_pretrained() methods in this repo load state dicts (plain dicts of tensors), weights_only=True is sufficient and doesn't change behavior for legitimate model files.

This is a narrow attack chain if the .pt files are only ever loaded from the official HuggingFace account, but in a shared research environment where files may be routinely passed around, a malicious file could propagate.

Affected call sites:

• dictionary.py: AutoEncoder, GatedAutoEncoder, JumpReluAutoEncoder, AutoEncoderNew from_pretrained()
• trainers/top_k.py: AutoEncoderTopK from_pretrained()
• trainers/batch_top_k.py: BatchTopKSAE from_pretrained()
• trainers/matryoshka_batch_top_k.py: MatryoshkaBatchTopKSAE from_pretrained()
• activault_s3_buffer.py: compile()

Test plan

• Existing pretrained models load correctly with weights_only=True
• pytest passes (no changes to logic, only to deserialization safety)