chore: drop dead hub-kernels workarounds

[samsja]

Summary

Follow-up to #3424 (native GPT-OSS implementation) and #3549 (Dockerfile --extra gpt-oss cleanup).

#3424 removed the gpt-oss extras group, whose only dependency was the HF hub kernels package. The package no longer appears in pyproject.toml or uv.lock — verified with uv lock --check (lock is current, no stale entry, so no lock update is needed). Without the package installed, transformers' is_kernels_available() is always False, so its hub-kernel integration can never engage.

This drops the two leftovers from that path:

• src/prime_rl/_compat.py — the lazy_load_kernel monkeypatch. The OfflineModeIsEnabled offline regression it worked around only triggers inside transformers' is_kernels_available() and _kernels_enabled guard (get_kernel() → HfApi), which is unreachable without the package. The ring_flash_attn shim stays; it is unrelated.
• src/prime_rl/trainer/model.py — the USE_HUB_KERNELS=NO env guard, which only mattered when installed hub kernels could intercept model modules.

Validation

• uv lock --check — passes, lock unchanged (no dependency change).
• ruff check / ruff format --check on both files — clean.
• grep confirms no remaining references to kernels hub package, lazy_load_kernel, USE_HUB_KERNELS, or OfflineModeIsEnabled in src/ or tests/.
• Note: also confirmed there is no any-depth/anydepth dependency anywhere in the repo or lockfile.

---

Note

Low Risk
Dead-code removal only; hub-kernel integration was already unreachable without the removed dependency.

Overview
Removes leftover Hugging Face hub-kernels compatibility code now that the kernels package is no longer a dependency (after the native GPT-OSS path and extras cleanup).

_compat.py: Deletes the lazy_load_kernel monkeypatch that caught OfflineModeIsEnabled for offline SLURM nodes. The ring_flash_attn shim for transformers ≥ 5.4 is unchanged.

trainer/model.py: Drops setting USE_HUB_KERNELS=NO at import time, which previously blocked hub-kernel module interception when the package was installed.

Behavior should be unchanged in current installs because transformers’ hub-kernel path never activates without the kernels package.

^{Reviewed by Cursor Bugbot for commit e52d13b. Bugbot is set up for automated code reviews on this repo. Configure here.}