Track upstream transformers async-load materialization crash workaround

[speediedan]

Summary

We currently carry a Windows-only workaround in src/interpretune/adapters/transformer_lens.py that sets HF_DEACTIVATE_ASYNC_LOAD=1 during Hugging Face model loading.

The workaround exists because transformers v5 async thread-based tensor materialization appears to trigger an upstream crash during from_pretrained() on our CI path.

Current local workaround

BaseITLensModule.hf_configured_model_init() now does the following on Windows when the env var is not already set:

• set HF_DEACTIVATE_ASYNC_LOAD=1
• call model_cls.from_pretrained(...)
• restore the previous environment afterward

Why track this

• The workaround is intentionally narrow, but it is still a behavioral fork from upstream defaults.
• We should reduce it to a documented compatibility shim, not a permanent hidden behavior.
• Once upstream behavior is fixed or clarified, we should remove the workaround and add a regression test covering the original failure mode.

Follow-up

• Capture the exact upstream exception/stack trace from the Windows CI failure if it is not already archived.
• Open or link the upstream transformers issue once we have a minimized repro.
• Remove the workaround after verifying the upstream fix in CI.